Mechanistic interpretability of cross-lingual concept representations in Tiny Aya — rise, peak, collapse.
-
Updated
Feb 23, 2026 - Python
Mechanistic interpretability of cross-lingual concept representations in Tiny Aya — rise, peak, collapse.
Early warning for LLM safety drift during fine-tuning: a Circuit Integrity Score read from the model's refusal circuit flags unsafe checkpoints before harm is measurable, with SafeTune repair and AuditKIT audit.
Find the harm, cut it out, keep it out: locating and removing harmful fine-tuning in Tiny Aya, keeping skills, surviving compression and later attacks. Lexsi AI Safety Hackathon, problems 9 to 12.
Medical fine-tuning degrades safety on Cohere Tiny Aya across 5 languages. However, a safety mixed SFT fixes this cutting attack success from 18.5% to 4.0% while maintaining full medical accuracy.
A small-model Gradio music app for era-to-era audio generation, fictional DJ narration, multilingual micro-lyrics, and safety-aware remix prompts.
One fix for a misaligned model family: measured cross-sibling repair of emergent misalignment in Cohere's Tiny Aya family (Lexsi AI Safety Hackathon, Track 2b)
Is safety repair language-fair? A pre-registered audit of multilingual refusal steering on Cohere's Tiny Aya, built on the Lexsi stack.
Answer or Abstain: an activation probe that makes Tiny Aya Global abstain on questions its passage cannot answer (SQuAD 2.0), measured beside a prompt instruction and a LoRA fine-tune. Lexsi AI Safety Hackathon 2026, Track 2b.
Hearth is an AI-integrated multilingual mobile communication and translation tool designed to bridge language barriers in women’s shelters.
Hinglish safety on Tiny Aya Global: refusals that drift out of the user's language, a PlainSFT LoRA repair (SafeTune) and before/after scoring (AuditKIT). Lexsi Labs hackathon, Track 2a.
Private, on-device translation across 67 languages on Apple Silicon — Cohere Labs Tiny Aya Global running locally with MLX and Streamlit.
Stops fine-tuning from breaking an LLM's safety: gates risky training data, watches refusal while it trains, rolls back before behaviour breaks. Built on Tiny Aya + Lexsi toolkits.
Does a multilingual bank chatbot keep required risk disclosures under 5 turns of pushback? LoRA SFT/DPO on Tiny Aya Fire, pre-registered. Lexsi AI Safety Hackathon 2026, Track 1.
Polyglot-MANSU: Cross-lingual machine unlearning and multilingual safety evaluation using Tiny Aya, with a focus on English, Hindi and Telugu.
Hearth is an AI-integrated multilingual mobile communication and translation tool designed to bridge language barriers in women’s shelters.
To associate your repository with the tiny-aya topic, visit your repo's landing page and select "manage topics."