A curated collection of tools, frameworks, and resources for AI-driven automated model training — letting AI agents autonomously run experiments, fine-tune models, optimize hyperparameters, and evolve themselves.
Inspired by Karpathy's AutoResearch, HuggingFace Skills, and the broader AutoML movement.
The paradigm is shifting: instead of manually tuning models, we now have tools that let AI agents design experiments, modify training code, evaluate results, and iterate autonomously — while you sleep.
This repository collects the best open-source tools and frameworks that make this possible across the full training lifecycle.
- Autonomous Experiment / Research Frameworks
- Agent-Driven Training Skills (HuggingFace Ecosystem)
- LLM Fine-Tuning Frameworks
- RL Alignment Training Frameworks (RLHF / GRPO)
- Automated Hyperparameter Optimization / AutoML
- Self-Evolving / Self-Play Training
- Synthetic Data Generation & Curation
- Knowledge Distillation
- Model Merging & Quantization
- Lightweight Pretraining & Distributed Training
- Inference Engines (for RL Training Loops)
- Multimodal Training Frameworks
- Experiment Tracking & Orchestration
- Benchmarks & Evaluation
- Coding Agents (for Training Script Development)
- Recommended Stacks
Core idea: AI agents autonomously design experiments, modify training code, evaluate results, and iterate. You sleep, AI experiments.
| Project | Description | Key Highlight |
|---|---|---|
| AutoResearch | AI agent runs autonomous ML experiments in a loop | 630 lines of Python, ~100 experiments overnight, 11% efficiency gain on GPT-2 training |
| AI Scientist v2 | Fully automated scientific discovery with agentic tree search | Hypothesis → Experiment → Paper, no human templates needed |
| AutoML-Agent | Multi-agent LLM framework for full-pipeline AutoML (ICML 2025) | Parallel specialized agents for preprocessing, architecture design, HPO; retrieval-augmented planning |
| auto-ml-agent | LLM-orchestrated autonomous ML pipeline | End-to-end: data preprocessing → model deployment, multi-agent architecture |
| MLAgentBench | Benchmark for evaluating AI agents on ML experimentation | 13 end-to-end ML tasks from CIFAR-10 to BabyLM |
| AutoAgent | Zero-code LLM agent framework with self-play customization | Create agents via natural language, iterative self-improvement |
| ShinkaEvolve | LLM-as-mutation-operator program evolution framework | Evolves programs for scientific discovery |
| AI-Supervisor | Autonomous research supervision via persistent Research World Model | Multi-agent consensus + Knowledge Graph; validates claims via GPU computation; self-correcting updates |
| ARIS | Lightweight Markdown-only skills for autonomous ML research overnight | Zero dependencies; cross-model review loops; 20+ GPU experiments per overnight run; works with any LLM agent |
| AutoKernel | Autoresearch for GPU kernels — sleep, wake to optimized Triton kernels | 1.2K+ stars; 5.29x speedup on RMSNorm H100; integrates with KernelBench |
| KernelAgent | Meta's autonomous deep-agent framework for GPU kernel generation | Under meta-pytorch org; auto-discover strategy; multi-model relay for kernel optimization |
| PaperOrchestra | Automated AI research-paper writer via coding-agent skills | Implements Google's PaperOrchestra; no API keys or LLM SDKs; works with Claude Code, Cursor, Aider |
| Deep Researcher Agent | Autonomous framework for 24/7 deep learning experimentation (U Tokyo, Apr 2026) | Leader-Worker multi-agent; Zero-Cost Monitoring; constant-size ~5K memory; 500+ cycles over 30 days at $0.08/24h |
| AI-Researcher | Fully autonomous research system: literature review → hypothesis → implementation → paper (NeurIPS 2025 Spotlight) | 4.8K+ stars; Resource Analyst agents reduce hallucination; Scientist-Bench benchmark for standardized evaluation |
| Deli_AutoResearch | Protocol framework for long-horizon (days–weeks) autonomous research — orchestrator + fresh-session workers, anti-loop & stall watchdog | By Chen Deli (@victorchen96); fresh-session architecture (no context accumulation), 3-layer heartbeat watchdog, JSON state persistence; produced 4 survey papers (265 pages, 1158 citations, 8.5+ self-score) incl. a 285B-param GRPO experiment in ~44h |
| AutoResearchClaw | "Chat an Idea. Get a Paper." Fully autonomous idea→paper pipeline (23 stages) | 13.5K+ stars, MIT; 6 human-in-the-loop modes (full-auto → co-pilot); 4-layer citation verification vs. hallucination; hardware-aware sandbox experiments; 8 papers across 8 domains; ARC-Bench |
| pi-autoresearch | Domain-agnostic port of Karpathy's loop — optimize any measurable metric (test speed, bundle size, Lighthouse, training loss) | 7K+ stars, MIT; extension = infra, skill = domain knowledge; append-only JSONL + markdown session log → resumes across restarts/context resets |
| OpenScience | Model-agnostic AI workbench that runs the full research loop — literature → hypothesis → code → experiment → analysis → write-up (Synthetic Sciences, Jul 2026) | Apache-2.0; 250+ editable skills + 30+ scientific databases (UniProt, PDB, ChEMBL, arXiv) as agent tools; any model swappable per-request (Claude/GPT/Gemini/GLM/DeepSeek/local); local-first, keys stay on your machine |
| Arbor | Generalist autonomous research agent built on Hypothesis-Tree Refinement (RUC NLPIR, arXiv 2606.11926) | Long-lived coordinator + short-lived executors in isolated worktrees; persistent tree binds hypothesis → artifact → evidence → insight so lessons propagate instead of scrolling off; 2.5x average relative held-out gain vs. Codex & Claude Code on the Autonomous Optimization benchmark; transfers to MLE-Bench Lite |
| Sibyl-AutoResearch | Self-evolving autonomous AI scientist built natively on Claude Code (arXiv 2605.22343) | "Autonomous research needs trial-and-error harnesses, not paper generators" — bounded trials preserve both positive and negative outcomes; 20+ specialized agents over a 19-stage pipeline; recurring process failures rewrite the harness itself |
| InternAgent-1.5 | Unified agentic framework for long-horizon autonomous scientific discovery (Shanghai AI Lab / InternScience, arXiv 2602.08990) | Three-subsystem loop — Generation (literature + hypothesis) → Verification (computation, simulation or wet lab) → Evolution (refine & propagate); persistent cross-session memory so failed directions aren't re-run; leads GAIA / HLE / GPQA / FrontierScience; covers both algorithm discovery and empirical discovery across physics, biology, earth & life science |
| MLEvolve | Self-evolving multi-agent system for end-to-end ML algorithm discovery and optimization (InternScience, arXiv 2606.06473) | Progressive MCGS extends tree search with cross-branch reference edges and an entropy-inspired explore→exploit schedule; Retrospective Memory pairs a cold-start domain knowledge base with dynamic global memory so experience is retrieved and reused across tasks; adaptive coding modes decouple strategic planning from code generation for stable long-horizon iteration; SOTA medal rate and valid-submission rate on MLE-Bench under a 12-hour budget (half the standard runtime), and beats AlphaEvolve on math algorithm optimization |
| The Agentic Scientist (AI Scientist v3) | Sakana AI's third-generation AI Scientist — orchestrates parallel off-the-shelf coding agents over the full research loop (Oct 2026, arXiv 2610.18933) | Apache-2.0; task → ideas → plan → sandboxed experiments → baselines & ablations → analysis → paper; a meta-controller enforces rubric checks per stage while workers share distilled conclusions through a common queue; dedicated Evidence Generation stage runs isolated controlled baselines/ablations; harness-swappable (Claude Code / Codex / Pi) — reportedly Kimi K3 + Pi matched Opus 4.7 + Claude Code on idea investigation at ~75% lower API cost |
"Vibe Training" — use natural language to drive the full model training lifecycle through coding agents.
| Project | Description | Key Highlight |
|---|---|---|
| HuggingFace Skills | Standardized ML skill packages for coding agents | 12 skills: model training (SFT/DPO/GRPO), vision training, experiment tracking, evaluation, dataset management |
| HuggingFace AutoTrain | No-code training platform | Upload data → auto model selection → training → evaluation → Hub publishing |
| ml-intern | Open-source ML engineer that reads papers, trains models, ships them | HF Hub + smolagents integration; 900+ stars; end-to-end LLM post-training on HF Jobs |
HF Skills covers:
hugging-face-model-trainer— Fine-tune LLMs with TRL (SFT, DPO, GRPO), 0.5B to 70B parametershugging-face-vision-trainer— Train object detection & image classification (RTDETRv2, YOLOS, ViT)hugging-face-jobs— Run compute jobs on HF infrastructure with cost estimationhugging-face-trackio— ML experiment tracking with real-time metricshugging-face-evaluation— Model evaluation with lightevalhugging-face-datasets— Dataset creation and management- Compatible with: Claude Code, OpenAI Codex, Google Gemini CLI, Cursor
The training engines. Upper-level agents (AutoResearch, HF Skills) ultimately call these frameworks to execute training.
| Project | Description | Key Highlight |
|---|---|---|
| Unsloth | Ultra-efficient LLM fine-tuning & RL | 2x faster, 70% less VRAM; custom CUDA kernels; MoE 12x faster; MCP Server available |
| Axolotl | Flexible, production-ready fine-tuning | YAML-driven; v0.8.x: QAT, sequence parallelism, GRPO, full RLHF pipeline |
| LlamaFactory | Unified fine-tuning with Web UI | LlamaBoard browser UI; 100+ models; SFT/RLHF/DPO/PPO |
| TRL | HuggingFace's RL training library | SFT, DPO, GRPO, PPO, KTO, ORPO; deep Transformers/PEFT integration |
| torchtune | PyTorch-native fine-tuning | No extra abstractions; multi-node support (Feb 2025) |
| NeMo AutoModel | NVIDIA's DTensor-native training library | Day-0 HuggingFace support; single-to-multi-node scaling |
| LMFlow | Extensible toolkit for fine-tuning large foundation models | LISA memory-efficient training (outperforms LoRA); FlashAttention; NAACL Best Demo Paper |
| H2O LLM Studio | No-code GUI framework for fine-tuning LLMs | Browser-based UI; LoRA/4-bit/8-bit; DPO/IPO/KTO; W&B integration |
| LitGPT | 20+ high-performance LLMs with pretrain/finetune/deploy recipes | CLI-driven; powered TinyLlama project; NeurIPS 2023 LLM Efficiency Challenge |
| InstructLab | IBM/Red Hat collaborative LLM customization via synthetic data | LAB alignment method; taxonomy-driven skill contributions; targets Granite models |
2025-2026 trend: GRPO (Group Relative Policy Optimization) is replacing PPO as the default alignment method — no critic model needed, simpler and more stable.
| Project | Description | Key Highlight |
|---|---|---|
| OpenRLHF | High-performance RLHF framework on Ray + vLLM | 70B+ full tuning; PPO/DAPO/REINFORCE++; async agent RLHF |
| verl | ByteDance's Volcano Engine RL for LLMs | GRPO/PPO in few lines; 3D-HybridEngine; used by ByteDance, Alibaba Qwen, UC Berkeley, LMSys |
| DAPO | Open-source RL system from ByteDance Seed + Tsinghua | 50 pts on AIME 2024 with Qwen2.5-32B; 4 key stability techniques; built on verl |
| AReaL | Fully asynchronous RL for LLM reasoning (Ant Group + Tsinghua) | 2.77x speedup vs synchronous; GSPO algorithm; Ascend NPU support |
| slime | LLM post-training framework for RL scaling (GLM team) | Powers GLM-4.5/4.6/4.7/5; Megatron + SGLang; RLVE (400 verifiable environments) |
| NeMo RL | NVIDIA's scalable post-training RL library | GRPO, SFT, DPO, DAPO; Ray-based; Megatron Core parallelism |
| NeMo Gym | Build RL environments for LLM training | Multi-step/multi-turn environments; interoperable with NeMo RL, OpenRLHF, TRL, Unsloth |
| rLLM | Post-training RL framework for language agents | Custom agents + environments → RL training → deployment; rLLM-FinQA-4B beats Qwen3-235B |
| RAGEN | Multi-turn RL framework for training reasoning agents | StarPO framework; 10 built-in environments; identifies "Echo Trap" instability |
| f-GRPO | f-Divergence based GRPO for general LLM alignment | KL/Reverse KL/Pearson/Hellinger/JS divergences; superior on both RLVR (math) and safety alignment; built on Unsloth |
| Tree-GRPO | Tree search for LLM agent RL (ICLR 2026) | 4x less rollout budget via shared prefixes; step-wise process supervision from outcome reward; tree-structured ReAct |
| SimpleRL-Reason | Simple RL recipe for reasoning (HKUST) | DeepSeek-R1-style; 7B achieves 33.3% AIME with only 8K examples; no SFT needed |
| SWE-RL | Meta's RL for software engineering reasoning | Llama3-SWE-RL-70B achieves 41% on SWE-bench Verified (NeurIPS 2025) |
| OpenManus-RL | RL tuning for LLM agents (UIUC + MetaGPT) | PPO-based; AgentGym environments + verl training |
| LlamaGym | Online RL fine-tuning for LLM agents | Define agent → create LLM → write RL loop |
| Reasoning Gym | Procedural reasoning environments for RLVR | 100+ tasks; NeurIPS 2025 Spotlight; unlimited controllable task generation |
| ART (Agent Reinforcement Trainer) | OpenPipe's open-source GRPO framework for training multi-step agents | Native multi-turn & tool-call support; vLLM inference + Unsloth training; LangGraph + MCP integration; auto LoRA checkpoint reload |
| MARTI | Tsinghua's framework for multi-agent LLM system RL training (MARTI-v2 Feb 2026) | Built on OpenRLHF; dynamic multi-agent workflows; GSPO loss; up to 32K token sequences; MARS2 tree-search RL for code |
| Arctic RL | Snowflake's open-source RL post-training backend (Jun 2026) | ZoRRo (Zero Redundancy Rollouts) auto-dedupes shared prompts — up to 6x actor-update & 3.5x end-to-end speedup; Arctic-Text2SQL-R2 from ~5 days → ~36h on 32 H200; plugs into existing RL frameworks |
| OpenRL | Google GKE Labs' self-hosted post-training / RL API (Jun 2026) | Tinker-compatible API; runs on any Kubernetes cluster; multi-job scheduling for high GPU utilization; imperative Python RL loops from your laptop; LoRA-first |
| OpenClaw-RL | Princeton's fully-async RL that trains agents from live conversations | Wraps model as OpenAI-compatible API; turns user corrections & trace errors into gradients; scalar PRM + token-level Hindsight-Guided On-Policy Distillation; zero manual labeling; personal + general agents |
| vime | Official vLLM-ecosystem RL post-training framework — Megatron training + vLLM rollout in one pipeline (Jun 2026) | Apache-2.0; built on slime's training design; fully async pipelines + train-inference mismatch correction; agentic RL for multi-turn tool calling & multi-agent; MoE/VLM support; AMD ROCm native |
| dLLM-RL (TraceRL) | First RL post-training framework for diffusion LLMs (ICLR 2026, Princeton/Gen-Verse) | Trajectory-aware RL + diffusion value model for stability; powers the SOTA TraDo series; TraDo-8B-Instruct beats Qwen2.5-7B on math; SFT/RL/RLHF across math/code/multimodal & architectures |
| Miles | Enterprise-grade RL post-training framework from the LMSYS / SGLang team (v0.1, Aug 2026) | Forked from and co-evolving with slime; SGLang rollout + Megatron-LM training at trillion-parameter MoE scale; fully async with configurable on/off-policy schedules; multi-LoRA adapters load straight into SGLang for rollout; agentic workflow (multi-turn, tool exec, sandbox, token-faithful trajectories); AMD ROCm support |
| Open-AgentRL | Gen-Verse's open agentic-RL stack — RLAnything + AutoTool (both ICML 2026) + DemyAgent | RLAnything generalizes one RL algorithm across terminal / GUI / SWE / tool-call settings (+9.1% OSWorld on Qwen3-VL-8B-Thinking, +18.7% AlfWorld on Qwen2.5-7B); AutoTool adds dynamic tool selection mid-trajectory; DemyAgent-4B matches 14B/32B agentic reasoning; the base OpenClaw-RL builds on |
| Agent-R1 | End-to-end RL for training tool-using LLM agents | Models each turn as a step-level MDP transition; StepPO (May 2026); built-in HotpotQA / ALFWorld / WebShop / paper-search environments |
| Agent Lightning | Microsoft's RL trainer that post-trains agents inside the harness you actually ship (v1.0, Aug 2026, arXiv 2608.17528) | MIT, 17.5K+ stars; ~3,500 lines of core Python; zero agent code changes — the agent reaches the model through an Agent Lightning proxy while tools, context, control flow and environments stay in the loop; verl + vLLM backend; native Kubernetes Jobs instead of an external sandbox service; 6K training samples took Qwen3.5-9B from 41.8% → 56.4% on SWE-bench Verified |
| AgentGym-RL | Modular framework for training LLM agents on long-horizon multi-turn decision making from scratch (ICLR 2026) | Fully decoupled environment / agent / training layers; PPO, GRPO, RLOO, REINFORCE++ plus complementary SFT / DPO / AgentEvol paradigms; five real-world scenario families (web navigation, deep search, digital games, embodied control, science tasks); ScalingInter-RL staged curriculum grows the interaction horizon from short to long for stable long-horizon training; trained agents match or surpass OpenAI o3 and Gemini-2.5-Pro across 27 tasks |
| FlashREINFORCE | Critic-free, single-rollout, asynchronous RL for agentic language models (Sep 2026) | One rollout per prompt — no GRPO group sampling, no PPO critic; batch-centered rewards + token-level importance sampling for staleness, a Sequence Trust Region (sequence-averaged Bernoulli KL proxy) to cap policy drift, and length-normalized Sample-Mean Optimization so long failed trajectories don't dominate; Qwen2.5-Math-1.5B averages 38.0 on five math benchmarks, +1.7 pts over GRPO with half the rollouts (256K vs 512K); available in OpenRLHF as a config (--algo.advantage.estimator flash_reinforce) |
| LEGO-RL | Harness-native online RL for coding agents — trains Claude Code, OpenHands and OpenCode on verl without changing their workflows (arXiv 2608.17393) | In-process LLM proxy captures raw generation streams for token-level alignment and trainer-side log-prob recompute; scalable sandbox orchestration with image caching and stage-wise reward-hacking defenses; built-in validation/monitoring plugin; SWE-bench Verified: OpenHands SDK 64.0% → 70.4%, Claude Code 62.4% → 68.2%, OpenCode 57.2% → 66.6%; code, models and datasets released |
| MCP-Universe RL | Salesforce's framework for training tool-use agents with RL, using MCP as the environment interface (arXiv 2608.22167) | Any tool already exposed as an MCP server plugs into training with no RL-specific glue; environment-orchestration layer provisions / isolates / recycles MCP environments over a pluggable container backend; staged rollout pipeline overlaps trajectories so GPUs stay busy while episodes wait on tools; backend-agnostic trainer with veRL and slime integrations; one config (only the task spec changes) improves reward for SWE, deep-research and general tool-use agents on gpt-oss-20b |
| LoongSage | Baidu BAIGE's production-grade, agent-native, fully asynchronous RL framework for frontier LLMs with validated recipes (arXiv 2610.16352, Oct 2026) | ~14K-line core on Ray + Megatron-Core + SGLang; AgentFlow orchestrates white-box and black-box agents — agents just call a standard chat-completion endpoint while multi-turn context, loss masking, trajectory collection and sandbox lifecycle are handled automatically; sliding-window staleness bounds + topology-aware TransferMesh weight sync; multi-teacher on-policy distillation via a Teacher Manager; four full training runs show steady reward gains with the train–inference gap kept under control |
| Project | Description | Key Highlight |
|---|---|---|
| AgentHPO | LLM-driven hyperparameter optimization | Matches/surpasses human best trials on 12 ML tasks with explainable results |
| AutoML-Agent | Multi-agent LLM framework for full-pipeline AutoML (ICML 2025) | Parallel specialized agents; retrieval-augmented planning; 14 datasets tested |
| Optuna | Industry-standard HPO framework | Bayesian search, pruning, distributed execution, visualization dashboard |
| Microsoft NNI | Full AutoML toolkit | Neural Architecture Search + HPO + model compression + feature engineering |
| W&B Sweeps | Automated hyperparameter search + tracking | Bayesian/Grid/Random search; Hyperband early stopping; cross-machine parallelism |
Core idea: Models generate their own training data to train themselves, reducing dependence on human annotations.
| Project | Description | Key Highlight |
|---|---|---|
| SPIN | Self-Play Fine-Tuning | Model plays against its previous iterations; outperforms DPO + GPT-4 preference data without extra annotations |
| SPPO | Self-Play Preference Optimization | Iterative policy updates approximating Nash equilibrium with convergence guarantees |
| SPC (Self-Play Critic) | Adversarial self-play for evolving reasoning critics | "Sneaky generator" vs "critic" game; eliminates manual step-level annotation |
| SPELL | Self-Play RL for Evolving Long-Context Language Models | Label-free self-play; base model surpasses instruction-tuned counterpart on long-context tasks |
| Multi-Agent Evolve | One LLM plays Proposer + Solver + Judge roles | Verified improvements on math, coding, reasoning with Qwen2.5-3B |
| Multiagent Finetuning | Multi-agent society from same base model | Multi-agent iteration keeps improving where single-model self-training plateaus |
| CORY | Cooperative multi-agent RL fine-tuning | Pioneer + Observer dual-agent paradigm (NeurIPS 2024) |
| AgentEvolver | ModelScope's self-evolving agent training framework | Unifies self-questioning + self-navigating + self-attributing; SeeUPO sequence-level agentic RL with convergence guarantees; generates its own training signal without human-labeled tasks |
Critical for automated training pipelines: generate high-quality training data at scale without manual annotation.
| Project | Description | Key Highlight |
|---|---|---|
| Distilabel | Framework for synthetic data and AI feedback pipelines | Modular pipeline; SFT/DPO/UltraFeedback techniques; any LLM provider |
| Magpie | Alignment data synthesis from scratch (ICLR 2025) | No prompt engineering needed; 4M instructions generated; matches Llama-3 Instruct |
| DataDreamer | Reproducible synthetic data generation (ACL 2024) | Multi-step prompting; generate/align/fine-tune/distill; built-in caching |
| Cosmopedia | Large-scale synthetic pretraining data pipeline | 25B tokens of synthetic textbooks/blogs; uses Mixtral-8x7B |
| InstructLab SDG | Synthetic data via LAB methodology (IBM/Red Hat) | Skills-SDG + Knowledge-SDG; minimal seed taxonomy → large-scale data |
| Persona Hub | Persona-driven synthetic data at billion scale (Tencent) | 1B diverse personas; 370M elite personas released |
| synth_gen | Execution-verified synthetic data (Meta) | Modular verifier system; parser-based verification for code |
| Evidently | Open-source synthetic data generation with user profiles | Model-agnostic; customizable personas & goals; no-code UI in Evidently Cloud; outputs to pandas DataFrame |
| NVIDIA Nemotron-4 340B | Open models for synthetic data generation pipeline | Base + Instruct + Reward models; commercial use allowed |
| NeMo Data Designer | NVIDIA's compound AI system for high-quality synthetic data generation | Dependency-aware fields; statistical samplers + LLM generation; LLM-as-judge scoring; image generation (v0.5.1); preview mode |
| synthetic-data-kit | Meta's CLI tool for generating fine-tuning datasets from any document (PDF, HTML, YouTube, DOCX) | 4-step pipeline: ingest → create → curate → save-as; QA pairs + CoT reasoning + summaries; LLM-as-judge filtering |
| DataFlow | LLM-driven framework for unified data preparation & workflow automation (OpenDCAI) | ~200 reusable LLM-based operators + 6 template pipelines (text, math reasoning, code, Text-to-SQL, agentic RAG, large-scale QA extraction); DataFlow-Agent turns a natural-language spec into an executable pipeline via operator synthesis + planning + iterative verification; drag-and-drop WebUI |
| DataArc-SynData-Toolkit | Unified closed-loop framework for multi-path, multimodal & multilingual data synthesis (DataArcTech + IDEA, arXiv 2605.08138) | 1.8K+ stars; config-driven end-to-end pipeline with both a visual UI and a simplified CLI; local synthesis / HuggingFace dataset retrieval / model distillation as interchangeable synthesis paths; quality-controllable loop runs through to downstream fine-tuning; strong low-resource-language coverage |
| Project | Description | Key Highlight |
|---|---|---|
| NeMo Curator | GPU-accelerated data preprocessing & curation | 30+ filters; fuzzy dedup 1.1T tokens in 1.8h on 64 A100s; 16x faster |
| DataTrove | Platform-agnostic data processing pipeline | Used for FineWeb and Cosmopedia; low memory; Slurm support |
| Dolma | High-performance dataset curation toolkit (AllenAI) | Built-in parallelism for billions of docs; used for OLMo (3T tokens) |
| Data Prep Kit | Unstructured data preparation (IBM) | Python/Ray/Spark runtimes; laptop to datacenter scaling |
Compress large models into smaller, deployable ones while preserving capabilities.
| Project | Description | Key Highlight |
|---|---|---|
| EasyDistill | Comprehensive distillation toolkit (Alibaba/ModelScope, EMNLP 2025) | Black-box + white-box KD; data synthesis + SFT + logits distillation + RL |
| DistillKit | Production-ready LLM distillation (Arcee AI) | Online and offline workflows; powers Arcee Virtuoso, SuperNova models |
| MiniPLM | Knowledge distillation for pre-training (Tsinghua, ICLR 2025) | Improved DPKD variant |
| DistiLLM | Streamlined distillation with contrastive approach (ICML 2024) | DistiLLM-2 contrastive distillation |
| EasyOPD | First unified on-policy distillation (OPD) framework — 10+ methods with one-line YAML switching (Jul 2026) | Built on verl; covers cross-tokenizer, self-distillation & step-wise OPD; separates config / supervision logic / execution; runnable configs + demo package |
| KDFlow | Decoupled LLM distillation framework — SGLang for the teacher, FSDP2 for the student (arXiv 2603.01875) | Fixes the systems mismatch of running a forward-only teacher inside the training backend; zero-copy hidden-state transfer + student-side logit recompute gives efficient full-vocabulary KD; off-policy / on-policy (OPD) / cross-tokenizer / multimodal / on-policy self-distillation, with multi-teacher, chunked loss and EMA teacher updates in v0.2.0; built-in Vanilla KD + DSKD and easy custom-algorithm registration; 1.4–6x faster than mainstream KD frameworks |
Combine multiple models or compress them for efficient deployment and training.
| Project | Description | Key Highlight |
|---|---|---|
| MergeKit | Leading toolkit for merging pretrained LLMs | SLERP, TIES, DARE, Passthrough, Evolutionary merge; works on CPU with 8GB VRAM |
| MergeLM | Language model merging codebase (ICML 2024) | Research-grade implementations |
| Project | Description | Key Highlight |
|---|---|---|
| GPTQModel | Production-ready LLM quantization toolkit | GPTQ, AWQ, QQQ, GPTAQ, EoRA, GAR; multi-backend CPU/GPU |
| AutoGPTQ | Easy-to-use GPTQ quantization | 8/4/3/2-bit; Marlin int4*fp16 kernel; ~150-200K monthly PyPI downloads |
| AutoRound | Advanced quantization via sign-gradient descent (Intel) | High accuracy at 2-4 bits; exports to GPTQ/AWQ/GGUF; broad HW compatibility |
| NVIDIA Model Optimizer | Unified quantization, pruning, distillation & speculative decoding | FP8/INT8/INT4; exports to TensorRT-LLM/vLLM; NeMo Megatron integration |
| TurboQuant | Google's KV cache compression (ICLR 2026) | 6x memory reduction at 3-bit with zero accuracy loss; PolarQuant + QJL; 8x perf on H100 |
| llama.cpp | LLM inference in C/C++ with GGUF quantization | Q4_K_M sweet spot: 92% quality, 75% size reduction; runs everywhere |
Pair these with autonomous experiment frameworks — fast, small-scale training is the foundation for autonomous experimentation.
| Project | Description | Key Highlight |
|---|---|---|
| nanochat | Minimal LLM training harness (AutoResearch's engine) | Single GPU; tokenization → pretrain → finetune → eval → chat; GPT-2 for ~$48 |
| Nanotron | Minimal 3D-parallel LLM pretraining | Data + Tensor + Pipeline parallelism; scales from experiments to production |
| Project | Description | Key Highlight |
|---|---|---|
| TorchTitan | PyTorch-native large-scale training platform | Up to 4D parallelism without model code changes; MXFP8 on Blackwell; elastic scaling |
| Open-dLLM | First open-source full stack for diffusion LLMs | Raw data → training → checkpoints → evaluation → inference, all-in-one |
Inference engines are critical for RL training — 80% of RLHF training time is spent on sample generation. Fast inference = fast training.
| Project | Description | Key Highlight |
|---|---|---|
| vLLM | Most mature open-source LLM serving engine | PagedAttention; 4x higher throughput on Blackwell; core engine for OpenRLHF |
| SGLang | High-performance serving for LLMs & multimodal | ~16,200 tok/sec on H100; RadixAttention; used by slime for RL training |
| TensorRT-LLM | NVIDIA's optimized inference library | FP8/FP4/INT4; EAGLE-3 speculative decoding; max GPU performance |
| LMDeploy | LLM compression, deployment & serving | TurboMind MXFP4; 1.5x vLLM performance; DeepSeek PD disaggregation |
| HuggingFace TGI | Multi-backend LLM serving (TensorRT-LLM, vLLM, llama.cpp) | Unified frontend; token streaming; HF Hub native; CPU/GPU/Inferentia support |
| NVIDIA Dynamo | Datacenter-scale distributed inference | 30x request throughput on DeepSeek-R1; disaggregated prefill/decode; Rust + Python |
Training models that understand text, images, video, and audio simultaneously.
| Project | Description | Key Highlight |
|---|---|---|
| LLaVA-OneVision-1.5 | Fully open-source multimodal training | Native-resolution images; SOTA performance; lower training costs |
| LLaVA-OneVision-1.5-RL | Democratized multimodal RL training | Open code, data, and models for multimodal RLHF |
| LLaVA-OneVision-2 | Next-generation fully-open multimodal training framework (EvolvingLMMs-Lab, arXiv 2605.25979) | Unified native OneVision-Encoder for image / long-video / spatial; 8B-class, no task-specific adapters; every stage, dataset & weight reproducible end-to-end |
| OpenRLHF-M | Multimodal model RLHF training | Extension of OpenRLHF for VLMs |
| LLaVA-KD | Multimodal knowledge distillation (ICCV 2025) | Distills large MLLMs into smaller ones |
| MoE-LLaVA | Mixture-of-Experts for vision-language models (TMM 2025) | Efficient multimodal MoE architecture |
| verl-omni | RL post-training for diffusion & omni-modality generative models (verl project) | Targets image/video/audio diffusion (Qwen-Image, Wan2.2, LTX-2.3), unified understanding+generation (BAGEL, HunyuanImage-3.0) and omni models (Qwen3-Omni); vLLM-Omni rollout with step-wise batching + FA3; DiffusionNFT & Diffusion DPO recipes; plugs into existing FSDP/USP parallelism |
| verl-vla | Unified vision-language-action (VLA) post-training framework (v0.1.0, Aug 2026) | One workflow for human-in-the-loop data collection → SFT → RL → policy evaluation; models, environments and algorithms integrate independently over a shared execution architecture; training workers, simulators and physical robots can run on separate nodes; teleoperate/intervene from any connected device |
| VLA Foundry | Toyota Research Institute's unified framework for training LLMs, VLMs and VLAs in a single codebase (arXiv 2604.19728) | Train an LLM → reuse the checkpoint to train a VLM → reuse again for a VLA, without switching codebases or adding external deps; most modules pure PyTorch; FSDP2 + WebDataset streaming; torchrun locally, AWS SageMaker/S3 on large clusters; pretrained LLM/VLM/VLA checkpoints released on the Hub |
| Project | Description | Key Highlight |
|---|---|---|
| Weights & Biases | Experiment tracking + sweeps + model registry | Industry standard; integrates with all major frameworks |
| MLflow 3.0 | Open-source experiment tracking + model serving | Self-hosted; nested experiments; model registry |
| ClearML | Open-source MLOps platform | 150K+ users at Fortune 500; auto-logging; pipeline orchestration; dataset versioning |
| HF Trackio | Lightweight experiment tracking in HF ecosystem | Deep integration with HF Skills; agents can read metrics and make decisions |
| RL-Insight | Online observability purpose-built for RL training loops (verl project) | One-command Prometheus + Tempo + Grafana stack; unifies actor, rollout, transfer-queue, engine and hardware metrics with RL state timelines in a single view while keeping training-side instrumentation lightweight; framework-agnostic data protocol; upstreamed into verl (v0.2.0, Jul 2026) and verl-omni |
| Benchmark | Description | Key Highlight |
|---|---|---|
| MLE-bench | 75 Kaggle ML engineering competition tasks | Evaluates AI agents on real ML engineering: training, data prep, experiments |
| MLAgentBench | 13 end-to-end ML experimentation tasks | Stanford SNAP; Claude v3 Opus best at 37.5% |
| PaperBench | Evaluates AI's ability to replicate ICML 2024 papers | 8,316 gradable tasks across 20 papers; best agent scores 21% |
| CORE-Bench | Computational Reproducibility Agent Benchmark | 270 tasks from 90 papers across CS, social science, medicine |
| MLRC-Bench | ML Research Competition challenges | Tests novel methodology development |
| AgentBench | Multi-dimensional benchmark for LLM agents | Tests across OS, database, knowledge graph, web, and game environments |
| SWE-bench Verified | Human-verified GitHub issue resolution | Industry standard for coding agents; top scores 70%+ |
| LiveBench | Monthly-updated contamination-free LLM benchmark | 6 categories (Math/Reasoning/Coding/Language/Data/IF); objective auto-scoring; no LLM judge needed |
| AgenticDataBench | Comprehensive benchmark for LLM data agents (Jul 2026) | 15 domains + 5 real-world B2B fintech use cases; fine-grained skill labels (schema inspection, joins, cleaning, viz, business-context reasoning); GitHub testbed + HF dataset, Apache-2.0 |
| Terminal-Bench 2.1 | 89 verified command-line agent tasks (Stanford + Laude Institute) | Covers software engineering, model training, security, data processing and sysadmin; Harbor task format + harness runs Claude Code, Codex CLI, OpenHands, Mini-SWE-Agent, Terminus 2; v2.1 patched 28 tasks so failures reflect agent capability, not environment gaps |
| AARR / AARRI-Bench | Benchmark suite scoring frontier LLMs and the agentic harnesses around them across the research lifecycle | Tests professionalism, thoroughness and nuanced scientific judgment in granular research scenarios instead of macro-level task completion; best configuration (mini-swe-agent + Claude Opus 4.7) reaches only 68.3%; finds minimalist scaffolds providing low-level primitives beat feature-heavy designs; probes field sensitivity and research ethics |
| SWE-bench Multimodal v2 | Visual software-engineering benchmark, fully open-sourced 2026-09-01 | 480 reproducible tasks retained from the original 517, with known flaky and ungradeable tests removed; issues carry screenshots of bugs, design mockups, diagrams and visual error context, so agents must act on image + text together; Docker environments rebuilt to fix dependency and browser drift; complete test split + evaluation tooling now runnable locally |
| AgenticBBO-Bench | Cross-domain benchmark for LLM agents as black-box optimizers, incl. hyperparameter optimization (NJU LAMDA, arXiv 2610.12183, Oct 2026) | Synthetic functions, HPO, database tuning, chip design and molecular design under one finite-budget protocol; agentic BBO beats direct LLM methods in all 5 domains and the best numerical optimizers in 4; ablations on tools, priors and LLM role; frontier challenge under the Codex harness puts GPT-6 Astra and DeepSeek-V4.1-Flash on the performance–cost Pareto front |
| Tool | Description | Key Highlight |
|---|---|---|
| DeepEval | Pytest-like LLM evaluation framework | v3.0: 14+ metrics; multi-turn simulation; DeepTeam for red teaming |
| Opik | Open-source LLM observability & evaluation (Comet) | Deep tracing; LLM-as-a-judge; hallucination detection; production dashboards |
| LMMs-Eval | Multimodal evaluation across text, image, video, audio | v0.6: eval-as-a-service; 7.5x throughput; 50+ tasks |
| Arize Phoenix | Open-source LLM observability and evaluation | Fully self-hosted; tracing, evaluation, retrieval analysis |
| LiveCodeBench | Contamination-free coding benchmark | Fresh problems from LeetCode/AtCoder/Codeforces |
These agents don't train models directly, but can write and debug training code, completing the automation loop when paired with HF Skills.
| Project | Description | Key Highlight |
|---|---|---|
| Aider | Terminal AI pair programming | Git integration; supports Claude/GPT/DeepSeek/local models |
| OpenHands | AI-driven software development (open-source Devin) | Autonomous code editing + execution + debugging; MIT license |
| SWE-agent | Autonomous GitHub issue fixer | SWE-bench open-source SOTA (NeurIPS 2024) |
| Open-SWE | LangChain's async cloud-hosted coding agent | Multi-agent (Planner + Reviewer); GitHub integration; auto PR creation |
| SERA | Ai2's open coding agent family | 54.2% on SWE-Bench; trains in 40 GPU-days (~$2K); all open |
| Cline | VS Code AI coding agent with 60K+ GitHub stars | MCP tool creation; 5M+ developers; human-in-the-loop approval; native subagents |
| OpenCode | Go-based terminal AI agent with 95K+ GitHub stars | Bubble Tea TUI; 75+ LLM providers; 6.5M monthly developers; SQLite persistence |
| Plandex | Terminal agent for large projects with 2M token context | Tree-sitter project maps; diff review sandbox; auto-debugging; 30+ languages |
| Roo Code | Terminal agent with 95K+ GitHub stars | 75+ LLM providers; plan-first development; 2.5M monthly developers |
| Gemini CLI | Google's open-source terminal AI agent (Apache 2.0) | Free tier: 60 req/min & 1K req/day; Gemini 2.5 Pro; MCP support; weekly stable releases |
| Claw Code | Open-source Rust rewrite of Claude Code agent harness architecture | 100K+ stars; fastest GitHub repo to 100K stars; multi-provider (Anthropic/OpenAI/xAI/Ollama/OpenRouter) |
| DeepSeek Harness | DeepSeek's official open-source agent harness — "everything is a plugin" (Aug 2026) | MIT, ~200K stars (100K in its first 2 days); model adapter, tool registry, session log, sandbox and the agent loop itself are all swappable plugins (Cordis-powered); Node.js, web-UI-first, runs fully local; developer preview |
| Live-SWE-agent | Runtime self-evolving software engineering agent | Expands and revises its own capabilities mid-task, no offline training; 79.2% on SWE-bench Verified with Claude Opus 4.5 |
| mini-swe-agent | 100-line minimalist coding agent (SWE-agent team, Princeton + Stanford) | >74% on SWE-bench Verified; no giant configs or monorepo; used by Meta / NVIDIA / IBM / Princeton / Stanford |
| Prime Agent | Prime Intellect's self-improving RLM harness — sub-agents are function calls inside a persistent IPython kernel (Aug 2026, arXiv 2608.23552) | MIT; the Recursive Language Model treats context as a variable and rlm(...) spawns a real child agent whose result returns into the session as data; Continual Harness keeps supplemental prompts, memories, skill descriptions and reusable subagent specs as durable state the agent refines through small evidence-backed updates; daemon-backed sessions, IPython state and schedules survive terminal detach; agent-to-agent messaging between live and retained subagents; built for long-running research evaluations; 95.5% on ARC-AGI-3 with Opus 5 |
HuggingFace Skills + Claude Code + Unsloth + W&B
Natural language → Claude Code orchestrates → HF Skills calls Unsloth for training → W&B tracks experiments.
AutoResearch + nanochat (single GPU)
Start before bed, wake up to ~100 autonomous experiment results.
Axolotl / LlamaFactory + OpenRLHF + Optuna + MLflow
YAML-configured training + automated HPO + full experiment tracking.
verl / OpenRLHF + vLLM/SGLang + Reasoning Gym + W&B
State-of-the-art RL training with fast inference engines and rich environments.
Distilabel / Magpie → Unsloth / TRL → DeepEval / LMMs-Eval
Generate data at scale → train efficiently → evaluate comprehensively.
- AutoResearch Paradigm: Karpathy proved "AI autonomously doing ML research" works with just 630 lines of code (66K+ stars by Apr 2026) — now a full ecosystem of derivatives: ARIS, AI-Supervisor, AutoResearchClaw (idea→paper), Deli_AutoResearch (long-horizon protocol), and the domain-agnostic pi-autoresearch port
- "Vibe Training": HF Skills enables natural-language-driven model training lifecycle
- GRPO Variants Proliferate: f-GRPO (f-divergence family), Tree-GRPO (tree search, ICLR 2026), DAPO — GRPO is the new default, and specialized variants are emerging fast
- RL Framework Explosion: verl, DAPO, AReaL, slime — every major lab now has an open-source RL training framework
- Self-Play Breakthrough: Multi-agent self-evolution (SPIN, MAE, SPC) overcomes single-model self-training plateaus
- Synthetic Data as Infrastructure: Distilabel, Magpie, Evidently make data generation a first-class pipeline stage; model collapse mitigation (Evol-Instruct) becoming standard
- MCP Standardization: Model Context Protocol adopted by OpenAI/Google/Microsoft as the "USB-C for AI agents"
- Single-GPU Research: Unsloth + nanochat + AutoResearch enables individual developers to do serious LLM research
- Inference-Training Convergence: vLLM/SGLang/TGI are now core components of RL training loops, not just serving
- Multimodal RL: LLaVA-OneVision-1.5-RL and OpenRLHF-M bring RL alignment to vision-language models
- Extreme Quantization: Google TurboQuant achieves 6x KV cache compression at zero accuracy loss (ICLR 2026); NVIDIA Model Optimizer unifies quantization/pruning/distillation
- Multi-Agent Coding Wave: Feb 2026 saw every major tool ship multi-agent capabilities (Grok Build, Windsurf, Claude Code, Codex CLI, Devin) — coding agents now routinely write training scripts
- Kernel-Level Autoresearch: Apr 2026 — autonomous agents for GPU kernel generation (Meta's KernelAgent, RightNow AI's AutoKernel) extend the "autoresearch" paradigm down to training infrastructure itself
- Multi-Agent RL Systems: MARTI, rLLM, ART, OpenManus-RL — RL training now targets multi-agent systems, not just individual models; centralized multi-agent interactions with distributed policy training becoming standard
- Agentic Data Creation: Meta synthetic-data-kit (document → dataset pipeline) and the Autodata framework show that AI agents iterating over their own data generation dramatically outperforms standard synthetic data methods
- Autoresearch Becomes a Primitive: Jun 2026 — the loop generalized beyond ML into a reusable "experiment → measure → keep/revert" infrastructure; pi-autoresearch optimizes any measurable metric (test speed, bundle size, Lighthouse), while idea→paper systems (AutoResearchClaw) and long-horizon protocols (Deli_AutoResearch) push autonomy from minutes to weeks
- RL Systems Layer + Continuous Online RL: Jul 2026 — RL efficiency is becoming its own infrastructure tier: Snowflake's Arctic RL (ZoRRo prompt-dedup, 3.5x end-to-end speedup) and Google's Tinker-compatible OpenRL (Kubernetes-native, multi-job GPU packing) decouple RL systems optimization from algorithms, while OpenClaw-RL turns live deployment conversations into always-on training signal — RL is shifting from offline batch jobs to standing, always-learning services
- On-Policy Distillation Goes Mainstream: Jul 2026 — OPD (student generates, teacher grades on-policy) is emerging as a standard post-training stage between SFT and RL; EasyOPD unifies 10+ OPD methods (cross-tokenizer, self-distillation, step-wise) on verl with one-line YAML switching, making it as configurable as fine-tuning
- RL Reaches New Substrates + Full-Science Autoresearch: RL post-training is spreading beyond autoregressive text — Gen-Verse's dLLM-RL/TraceRL brings trajectory-aware RL to diffusion LLMs (SOTA TraDo series), and the vLLM project's own vime standardizes RL post-training inside the inference-engine ecosystem; meanwhile the autoresearch loop generalizes from ML to full-science workbenches (OpenScience: literature → hypothesis → experiment → write-up across ML/bio/physics/chem)
- RL Frameworks Fork into a Per-Modality Matrix + Enterprise RL: Aug 2026 — the verl ecosystem split into three specialized lines (verl for text, verl-omni for diffusion/omni generative models, verl-vla for vision-language-action robot policies), so RL post-training now covers image/video/audio generation and physical robot control, not just text; in parallel LMSYS/SGLang's Miles pushes RL into trillion-parameter MoE production runs where precision, stability and observability matter as much as throughput
- Autonomous Research Shifts from Producing Papers to Accumulating Knowledge: The newest research agents attack the same failure — that each experiment starts from scratch. Arbor's Hypothesis-Tree Refinement persists hypothesis → artifact → evidence → insight in a tree and propagates lessons upward (2.5x the average relative held-out gain of Codex/Claude Code at equal budget), while Sibyl argues autonomous research needs self-evolving trial-and-error harnesses, not paper generators — recurring process failures rewrite the harness itself
- The Agent Harness Becomes Swappable Infrastructure: DeepSeek open-sourced its own harness (100K stars in 2 days) where the model adapter, tool registry, sandbox and the agent loop itself are plugins — following Claw Code's clean-room rewrite, the harness layer is now commodity open infrastructure rather than vendor lock-in; Terminal-Bench 2.1 (Stanford + Laude, 89 tasks incl. model training) is emerging as the shared measuring stick across harnesses
- Self-Evolution Moves to Runtime: The boundary between training-time and inference-time improvement is collapsing. Live-SWE-agent expands and revises its own tools and workflow mid-task with no offline training (79.2% on SWE-bench Verified), while AgentEvolver self-evolves during training via a self-questioning / self-navigating / self-attributing loop that generates its own reward signal without human-labeled tasks. On the training side, Agent-R1 makes every tool call a step-level MDP transition, folding tool use, memory and credit assignment into one trainable substrate — so agentic RL now targets multi-turn tool-using policies directly, not single-response ones. mini-swe-agent is the counterweight: 100 lines beats >74% on SWE-bench Verified, showing scaffold complexity is not where the capability lives
- Training Moves Inside the Harness You Ship: Sep 2026 — the newest frameworks stop asking you to rebuild your system for training. Microsoft's Agent Lightning v1.0 post-trains an agent in its deployment harness: tools, context, control flow and environments stay in the loop while the agent reaches the model through a drop-in proxy with zero code changes (~3,500 lines of core Python on verl + vLLM, Kubernetes Jobs as the sandbox; 6K samples → +14.6 pts on SWE-bench Verified). The same train/serve collapse shows up across the stack — VLA Foundry keeps LLM → VLM → VLA pretraining in one codebase instead of three handoffs, InternAgent-1.5 gives the discovery loop persistent cross-session memory so failed directions aren't re-run, and DataArc-SynData-Toolkit closes synthesis → fine-tuning into a single quality-controlled loop. Evaluation is following suit: AARRI-Bench scores the harness, not just the model, and still finds the best configuration at only 68.3%
- Research Loops Grow a Memory; the RL Stack Grows an Ops Tier: Sep 2026 — two complementary maturations. On the research side, the fix for "every experiment starts from scratch" is now concrete machinery: MLEvolve replaces one-shot tree search with Progressive MCGS (cross-branch reference edges, entropy-scheduled explore→exploit) plus a Retrospective Memory that retrieves prior task experience — reaching SOTA on MLE-Bench at half the standard 24h budget and beating AlphaEvolve at math algorithm optimization. Prime Agent pushes the same idea into the harness itself: its Continual Harness treats prompts, skills and reusable subagent specs as durable state the agent edits from evidence, while sub-agents become literal function calls in a persistent REPL. On the systems side, RL is acquiring the operational tier ordinary training has had for years — RL-Insight gives rollout / actor / transfer-queue telemetry a Grafana-grade home (now upstream in verl), KDFlow splits the distillation pipeline along its real seam (SGLang for the forward-only teacher, FSDP2 for the student) for 1.4–6x speedups, and AgentGym-RL turns long-horizon stability into a configurable curriculum (ScalingInter-RL) instead of a tuning art. Evaluation keeps pace: SWE-bench Multimodal v2 opened its full 480-task split for local reproducible runs
- Agentic RL Standardizes the Environment Interface and Cheapens the Rollout: Oct 2026 — the bottleneck in agent RL is shifting from algorithms to environment-integration cost and sampling cost. Salesforce's MCP-Universe RL treats MCP itself as the RL environment interface: any MCP server plugs in with zero integration code, and one config (only the task spec changes) trains SWE, deep-research and general tool-use agents on gpt-oss-20b. LEGO-RL wires three off-the-shelf harnesses — Claude Code, OpenHands, OpenCode — into verl unchanged and lifts each by 6–9 pts on SWE-bench Verified, so after Agent Lightning, "train inside the native harness" is becoming the default for coding-agent RL. On the sampling side, FlashREINFORCE drops both GRPO's group sampling and PPO's critic — one trajectory per prompt, stabilized by a sequence trust region and sample-mean optimization — beating a GRPO baseline with half the rollouts, and ships in OpenRLHF as a one-line config
- Research Agents Become an Orchestration Layer over Coding Agents; Agents Take Over Tuning: Oct 2026 — Sakana's third-generation AI Scientist, The Agentic Scientist, no longer builds a monolithic agent: a meta-controller orchestrates parallel off-the-shelf coding agents (Claude Code / Codex / Pi) and a separate Evidence Generation stage has isolated agents run baselines and ablations — the value of a research system is moving from the agent itself to orchestration and evidence gating, with the harness as a swappable part. On the training side, Baidu's LoongSage lets black-box agents enter a fully asynchronous RL loop through nothing but a chat-completion endpoint (AgentFlow), continuing the zero-change-integration trend. On evaluation, AgenticBBO-Bench shows LLM agents now beat the best numerical optimizers in 4 of 5 black-box optimization domains including HPO — the "tuner" role in AutoML is being taken over by agents
- Awesome LLM Synthetic Data
- Awesome Knowledge Distillation of LLMs
- Awesome Model Merging
- Awesome LLM Quantization
- Awesome LLM Inference Engine
- LLM Datasets
- LLM Distillation Playbook
Contributions are welcome! Please open an issue or submit a PR if you know of tools that fit this collection.
Criteria for inclusion:
- Must be directly usable for automated model training workflows
- Preference for open-source projects with active maintenance
- Focus on tools that leverage AI/LLMs to automate the training process itself
This curated list is released under CC0 1.0.
Compiled March 2026, updated 2026-10-16. Project statuses may change — check individual GitHub repos for the latest.