ludocomito/qwen2.5-3b-tribe-dpo-lora

Qwen2.5-3B-Instruct TRIBE DPO LoRA

This repository contains a PEFT LoRA adapter trained from Qwen/Qwen2.5-3B-Instruct for poetry / creative-writing generation using a TRIBE-derived reward signal.

The adapter is the best checkpoint from the DPO run selected by lowest eval loss.

Training Summary

• Base model: Qwen/Qwen2.5-3B-Instruct
• Method: SFT warmup followed by DPO
• Reward source: pooled TRIBE poetry-vs-control activation axis, adjusted for length and repetition
• DPO pairs: 592 total
• Train pairs: 535
• Eval pairs: 57
• Best checkpoint: checkpoint-50
• Best eval loss: 0.6571338772773743
• Train loss: 0.6417140785385581
• Final saved adapter: models/dpo_lora_best

The DPO preference pairs were built from generated candidates scored by the TRIBE-derived reward. Higher-scoring candidates were used as chosen responses and lower-scoring candidates as rejected responses, with a minimum adjusted reward margin of 0.12.

Intended Use

This is an experimental research adapter for probing whether a brain-model-derived reward can steer a small instruction model toward poetic / affective text. It is not a general-purpose writing model and should be treated as a research artifact.

Example loading:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-3B-Instruct"
adapter_id = "ludocomito/qwen2.5-3b-tribe-dpo-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

Included Artifacts

• LoRA adapter weights and config
• tokenizer files copied from the training output
• DPO training summary
• DPO dataset summary
• step-level DPO metrics
• probe generations sampled over training
• selected TRIBE pooled-axis features used for scoring

Limitations

• The reward is not a human preference model. It is a pooled TRIBE activation axis trained to separate poems from controls in the earlier validation experiments.
• The eval split is small, with 57 preference pairs, so metrics should be read directionally.
• The model may learn reward-specific texture rather than robust literary quality.
• This adapter should be evaluated with fresh prompts and manual inspection before any broader use.

Local Experiment Context

The run used 1,024 scored generated candidates, balanced as 4 candidates for each of 256 prompts. The full 2,048 generated candidates were preserved locally, but only the balanced subset was scored by TRIBE for the final DPO run.