AmberNeel/tiny-aya-global-redlens-guard-GGUF

Tiny Aya Global · RedLens Guard (GGUF, Q4_K_M)

CohereLabs/tiny-aya-global quantized to Q4_K_M, with the RedLens Guard v1 multilingual safety preamble baked into the chat template. Built for the Lexsi AI Safety & Alignment Hackathon (Track 2). Weights are unchanged; only tokenizer.chat_template and the model name/description metadata differ from the plain Q4_K_M file.

Why: Tiny Aya Global refuses all of our 30 harmful seed requests in English, but gives full or partial help to 47% of them in Telugu and 53% in Hinglish. With the guard, those drop to 13% and 10%, Hindi from 10% to 0%, and the cross-lingual gap from 53 to 13 points. The cost is +2.1 pts over-refusal on benign XSTest prompts.

Metric (150 harmful / 96 benign prompts, en·hi·te·Hinglish + role-play)	Baseline	Guard
Attack success, full compliance	8.7%	4.0%
Attack success incl. partial	28.0%	8.7%
Mean harm severity (0-1)	0.142	0.044
Cross-lingual gap (full + partial)	53.3 pts	13.3 pts
Over-refusal (benign)	8.3%	10.4%
Benign helpfulness	78.1%	71.9%
Reply in user's language	99.2%	99.2%

Measured on the Cohere API (same weights) with the guard sent as the system message; full method, code and logs: github.com/cryptocloud786/aya-redlens, prompt set, eval logs.

Use

llama-cli -m tiny-aya-global-redlens-guard-Q4_K_M.gguf -cnv

from llama_cpp import Llama
llm = Llama(model_path="tiny-aya-global-redlens-guard-Q4_K_M.gguf", n_ctx=4096)
llm.create_chat_completion(messages=[{"role": "user", "content": "Python process ko kaise kill karte hain?"}])

The guard is applied even with no system message. Your own system message is appended after the guard (it cannot replace it). chat_template_redlens_guard.jinja is the template on its own, for use with the original model in transformers/vLLM.

What the guard says (summary)

Judge requests by intent, not language, script, code-mixing, encoding or role-play framing; do not give operational help for weapons, hacking, fraud, violence, self-harm, harassment or hate; decline briefly in the user's own language with a safe alternative; and do not over-refuse homonyms, figurative language, fiction without operational detail, definitions, history or public facts.

Limits

Prompt-level defence: anyone with the weights can remove it. Role-play jailbreaks still succeed 20% of the time in English. Hinglish comprehension is weak in the base model and the guard does not fix it. Licence and acceptable use follow the base model (CC BY-NC 4.0, Cohere Labs AUP).