tiny-aya-fire-lexsi-ctheta-repair
surgical safety repair without fine-tuning
2 min readupdated
tiny-aya-fire-lexsi-ctheta-repair is a patched version of the compact multilingual language model CohereLabs/tiny-aya-fire (3.35B parameters, 36 layers, Cohere2ForCausalLM architecture), tailored for South Asian languages and English. The lexsihack team fixed multilingual safety vulnerabilities and leaks of malicious cybersecurity instructions using Circuit-Restricted Weight Arithmetic (C-ΔΘ)—an offline weight-editing method applied inside a targeted neural circuit. Instead of standard fine-tuning (SFT or DPO), which often makes models refuse even harmless work-related prompts, they used gradient attribution (EAP-IG) to pinpoint 11 key MLP matrices and adjusted only those. The model is a solid fit for developers building enterprise assistants and SecOps tools who need a local LLM without the added latency of external guardrails.
Key features
- Surgical precision: only 11 out of 36
mlp.down_proj.weightmatrices (layers 0, 1, 2, 6, 13, 23, 27, 28, 29, 33, 35) were modified, leaving the rest of the model weights bit-for-bit untouched. - Fast checkpoint creation: circuit discovery and weight adjustment took around 17 seconds on a single NVIDIA L4 GPU with no full retraining required.
- Fewer cybersecurity leaks: attacker-useful responses dropped from 9.4% to 5.0%–5.5% while preserving 10 out of 10 practical defensive SecOps scenarios.
- Stronger multilingual safety: refusals on harmful prompts in Bengali and Banglish (
BanglaSafe) jumped from 45.0% to 64.0%, with no increase in false refusals on benign prompts.
What you can do
- Blue Team assistant: generate SIEM/Sysmon detection rules, configure
fail2banandnginx, or triage incidents without risking ready-to-use exploit outputs. - Local multilingual bots: deploy a lightweight 3.35B model in an air-gapped environment supporting English, Hindi, Bengali, Tamil, Telugu, and Kannada without extra input/output safety classifiers.
- Interpretability research: reproduce the
CuratorKIT → CircuitKIT → SafeTune → AuditKITpipeline from the authors' repo to surgically steer other compact LLMs.
Example
You can test the model on defensive and borderline prompts (such as writing fail2ban rules, configuring nginx, or working in local Indian languages) in the authors' interactive Gradio and Streamlit demo app (hf_space/app.py in the RealASHWIN122/lexsihack repository).
Read more
DetailsBase model: CohereLabs/tiny-aya-fire (3.35B, bfloat16)
Benchmarks: HarmBench — 71.0% refusals (+2.0 pp), StrongREJECT — 96.5%, BanglaSafe — 64.0% (+19.0 pp), OR-Bench-Hard-1k — 58.2% false refusals (-0.5 pp)
Knowledge retention: WMDP-Cyber 39.6%, WMDP-Bio 61.5%, WMDP-Chem 45.3% (under 0.7 pp shift)
Compute: ~17 seconds on NVIDIA L4, repair coefficient α = 0.50










Comments