kelvinyelyen/tiny-aya-global-hedge-lora

tiny-aya-global Hedge-Instruction LoRA

A small LoRA adapter fine-tuned on top of CohereLabs/tiny-aya-global to fix one specific, verified failure mode: the base model ignoring an explicit "if unsure, say I'm not sure" instruction and answering confidently instead.

This adapter was built as a follow-up to a qualitative blind-spot evaluation of the base model: kelvinyelyen/tiny-aya-global-blindspots. One finding there was that, given a prompt that explicitly instructed the model to hedge if uncertain, the base model answered a stale fact (a former president of Ghana) with full confidence, ignoring the instruction outright. This adapter targets that specific behavior.

Method

• Base model: CohereLabs/tiny-aya-global, loaded in 4-bit (bitsandbytes, NF4).
• Adapter: LoRA, rank 8, alpha 16, applied to: k_proj, o_proj, q_proj, v_proj (detected at runtime from the base model's actual module names, not assumed from a different model family).
• Training data: 15 hand-built examples (10 pairing an explicit hedge instruction with a response that actually hedges, 5 plain factual control questions with confident correct answers). The control examples exist so the adapter learns selective hedging tied to genuine uncertainty, not "always hedge" regardless of the question.
• Training: 3 epochs, batch size 2 (grad accumulation 4), learning rate 0.0002.

Results

(Fill in from the before_after list printed in Section 8 after running the notebook — not pre-filled here, since these are actual model outputs that need to be observed, not assumed.)

Prompt	Before	After
...	...	...

Also report here whether the adapter hedged appropriately on the plain factual control prompts (it shouldn't) — that's the real test of whether this generalized rather than overfit to "always say I'm not sure."

Limitations

• Trained on 15 examples — small enough that memorization of the exact training phrasing is a real risk. Check the training loss curve and the control-prompt behavior before trusting this as a general fix.
• Targets one failure mode only (ignored hedge instructions). Says nothing about the model's other documented blind spots (exact counting, low-resource translation, formal logical inference) — see the blind-spot dataset above for those.
• Single training run, no held-out validation set beyond the 5 qualitative examples in Section 8.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", device_map="auto")
model = PeftModel.from_pretrained(base, "kelvinyelyen/tiny-aya-global-hedge-lora")
tokenizer = AutoTokenizer.from_pretrained("kelvinyelyen/tiny-aya-global-hedge-lora")