Tiny Aya Global — safety-mixed medical SFT (recommended)

Plain medical fine-tuning made Tiny Aya less safe in every language I tested. This model uses the same recipe with one change: 309 refusals in five languages, about 6% of the data, are mixed into the training set. The original model wrote those refusals itself. The result is an attack success rate of 4.0%, lower than the Tiny Aya Global I started from (10.5%), with medical accuracy unchanged.

This is a QLoRA adapter for CohereLabs/tiny-aya-global, trained with AlignTune. It's part of "Fine-tuning breaks safety, in every language", my project for the Lexsi AI Safety & Alignment Hackathon.

Results

Metric (%) Tiny Aya Global This model
Medical accuracy (Global-MMLU, 5 langs) ↑ 52.0 52.8
General accuracy ↑ 50.8 47.1
Harmful ASR, aya_redteaming (Qwen3Guard) ↓ 10.5 4.0
Harmful ASR (Llama-Guard-3-1B) ↓ 14.8 4.5
Harmful refusal ↑ 58.5 71.5
XSTest unsafe ASR ↓ 1.0 0.0
XSTest over-refusal ↓ 6.8 9.6
Benign-medical over-refusal ↓ 0.8 3.6

Training

  • Data: 4,497 medical Q&A pairs plus 309 refusals. Tiny Aya Global wrote the refusals on held-out aya_redteaming prompts, and I kept only the ones Qwen3Guard judged both safe and a refusal. Curated with CuratorKIT (safety_mix config).
  • QLoRA NF4 + double quant, bf16 compute; LoRA r=16, alpha=32, dropout 0.05 on q/k/v/o/gate/up/down projections
  • lr 2e-4, 1 epoch, effective batch 16, max length 640, paged AdamW 8-bit, gradient checkpointing, seed 42
  • The tied 262k-token embedding is kept in bf16. PEFT would otherwise up-cast it to fp32, which runs out of memory on 8 GB.
  • Trained on an A40 (48 GB) using free RunPod credits; the same recipe fits in 8 GB

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "KiranPal18/tiny-aya-global-med-safemix")
stop = tok.convert_tokens_to_ids(["<EOS_TOKEN>", "<|END_RESPONSE|>", "<|END_OF_TURN_TOKEN|>"])
msgs = [{"role": "user", "content": "What are common symptoms of iron-deficiency anaemia?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**ids, max_new_tokens=256, do_sample=False, eos_token_id=stop)
print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))

How it was evaluated

Every model was evaluated the same way: loaded in 4-bit NF4, greedy decoding, Tiny Aya's own chat template, in English, Hindi, Spanish, French and Arabic. Generation stops on <EOS_TOKEN>, <|END_RESPONSE|> and <|END_OF_TURN_TOKEN|>. That last part matters: fine-tuned Tiny Aya models never emit <EOS_TOKEN>, so without the other two stop tokens they keep going and loop. Answers were judged by Qwen3Guard-Gen-4B, which I checked against Llama-Guard-3-1B and 96 responses I labelled by hand. All prompts, generations and labels are in KiranPal18/aya-med-safety-evals.

What it's for, and what it isn't

This is a research artifact for studying how a domain fine-tune changes safety behaviour across languages. Please don't use it for medical advice. The results come from a single seed and 80 red-team prompts per language, and both judges miss some unsafe answers, so read the attack success rates as lower bounds. Licence: CC-BY-NC-4.0, inherited from Tiny Aya. Full write-up and code: https://github.com/KiranPal18/MultiLingual-Safety-Drift

Downloads last month
36
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for KiranPal18/tiny-aya-global-med-safemix

Adapter
(395)
this model

Dataset used to train KiranPal18/tiny-aya-global-med-safemix

Space using KiranPal18/tiny-aya-global-med-safemix 1

Collection including KiranPal18/tiny-aya-global-med-safemix