Mizore66/tiny-aya-global-evidencetune-lora

Tiny Aya Global — EvidenceTune LoRA

A LoRA adapter (r = 16; attention and MLP projections) for CohereLabs/tiny-aya-global. It teaches the model to change its answer only when a correction cites evidence that settles the question.

When a user pushes back, the model replies with:

ACTION: UPDATE | CLARIFY | RETAIN
ANSWER: yes | no
<one sentence>

Training

• Trained with SafeTune PlainSFTTrainer: 2 epochs, LR 1e-4, batch 4, seed 0.
• Data: the heldout_templates/train split of Mizore66/evidencetune-bench. That's 1,060 ContractNLI rows from 12 statement templates plus 2,120 matched VitaminC rows, prepared with CuratorKIT.
• lexsi_provenance.json records the full lineage.

Evaluation

Frozen set: 136 fresh cases from 5 statement templates never seen in training. Evaluated with AuditKIT, greedy decoding.

	this adapter	contracts-only adapter	word-overlap rule	base
matched valid/unsupported pairs handled	115/136	102/136	95/136	~0
lexical traps (of 41)	31	27	0	~0
bare pressure resisted	136/136	136/136		0%
unwarranted answer changes (of 272)	29	35		~all

Safety retention: on 200 public multilingual aya_redteaming prompts, judged by Aya Expanse 32B, harmful replies were 6 for an adapter trained on the in-domain data vs 10 for base.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global"),
    "Mizore66/tiny-aya-global-evidencetune-lora",
)

Prompt with the instruction and input format from the dataset rows, as a single user turn through the chat template.

Limitations

• One seed; English task data; labels are dataset proxies.
• Not compared with prompting-only baselines yet.

Licence: CC BY-NC 4.0, following the base model.