retroib/qwen2.5-14b-exiobase-lora

Qwen2.5-14B-Instruct - EXIOBASE Activity Mapper (LoRA)

LoRA adapter mapping free-text procurement items to one of 83 EXIOBASE emission-factor activities, returning JSON: {activity, reason, confidence_level}.

Results (held-out)

On the unleaked silver-51 test, fine-tuning adds +17.7 exact points over the in-context base and +27.5 over the embedding baseline. (*gold rule-book scores are optimistic - the rule book was partly derived from the gold items.)

Training

Base Qwen2.5-14B-Instruct (4-bit). LoRA via Unsloth + TRL, r=16 a=16, 2 epochs, lr 2e-4. 498 silver labels from a Qwen2.5-32B teacher, cross-checked by Claude + Gemini, frozen disjoint split. Teacher and student both open-source.

Limitations

The reason field was trained on a placeholder and is not informative. Coverage spans 56 of 83 activities; long-tail classes are weak.

Code, notebook and writeup: https://github.com/ibrahim-2337/carbonsifr-exiobase-mapping