lora_gemma4-e4b-prestashop

QLoRA adapter for Gemma 4 E4B (google/gemma-4-E4B-it) trained to follow a fixed bug-fixing flow on PrestaShop (PHP): localise (keywords) → read (files) → edit (SEARCH/REPLACE blocks). Part of the Kaggle Gemma 4 Developer Agent Paper Track submission — paper and code: https://github.com/ba-rem26007/gemma4-legacy-replay (docs/KAGGLE_FINAL_WRITEUP.md).

Training data

585 prepared examples, all from PrestaShop bugs merged before 2025-06-01 (Gemma 4's declared cutoff is January 2025):

  • 569 paths reconstructed deterministically from official upstream fixes (trajectories/reconstruct.py);
  • 16 paths produced by Gemma 4 31B itself and verified by execution (trajectories/self_paths.py).

Examples touching any function also fixed after the split were excluded. No proprietary-model output is included. With MAX_LEN = 2048, 89 examples were actually used. Exact snapshot: training/snapshots/trajectories_v15.tar.gz.

Training

4-bit QLoRA, r = 16, α = 32, 3 epochs (18 steps), lr 5e-5, 2× T4 on Kaggle, 7,209 s, mean training loss 1.192 (training/snapshots/train_kaggle_v15.py, log training/lora_final/gemma-4-qlora-training-prestashop.log). A chunked cross-entropy (training/chunked_loss.py) avoids materialising the full logits tensor over the 262k vocabulary. adapter_model.safetensors sha256 starts with fac3f1af8b0fb855.

Evaluation (33 hidden-oracle TEST bugs, PrestaShop 9.1.x, merged after the split)

Condition E (rules + glossary + replay spec, 2 retries), 3 runs per model on the same infrastructure (base loaded in 4-bit NF4 as in training; docs/RESULTATS_E4B.md, recomputed by notebook/verification.ipynb):

Model Solved per run Mean Right file read Δ vs base (95% CI)
Gemma 4 E4B base (no adapter) 3, 4, 4 3.67 / 33 13–16 —
E4B + this adapter (v15, 89 examples) 1, 1, 2 1.33 / 33 11–13 −7.1 pts [−16.2, 0.0]
E4B + v16 adapter (453 examples, ≤ 4096 tokens) 0, 0, 0 0 / 33 11 −11.1 pts [−23.2, −2.0]

This adapter does not improve Gemma 4 E4B on this benchmark; the larger v16 adapter degrades it. Patch application rates are unchanged; localisation drops. An earlier single-run figure (4/33, 28 Sept.) is within the base model's noise. We publish the adapter for reproducibility of this negative result.

Limitations

Small training set (89 examples), n = 33 test bugs, single run for E. PrestaShop is © PrestaShop SA and contributors (OSL-3.0).

Downloads last month
10
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for elrems/lora_gemma4-4b-prestashop-v1

Adapter
(508)
this model