Modelpublic

singhabhishekkk/apprentice-gemma4-e4b-lora-cuad-mi300x

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes3downloads

Model Card

Apprentice Gemma 4 E4B LoRA, trained on AMD MI300X (contract clause extraction)

Built with Gemma. bf16 LoRA (r=16, alpha 16, 3 epochs, TRL + PEFT) trained on an AMD MI300X with ROCm for the AMD Developer Hackathon: ACT II. 140 golden CUAD rows, evaluated on the same 60 held-out rows as the public benchmark.

Results (measured 2026-07-10, field-level F1)

System	Score
Gemma 4 E4B raw	36.33
Gemma 4 E4B fine-tuned (this adapter)	61.67
gpt-5.4-mini, GEPA-optimized (best published teacher)	36.33

Train wall time 505.8 s (~8.4 GPU-minutes). Full run notebook and report: github.com/singhabhishekkk/apprentice-amd-hackathon

Caveat: 60-row held-out eval with exact-string field matching. Re-validate on your own data before production use.