Gemma-4 E4B LoRA — SCIDOCA 2025 Task 1 (citation discovery) — pointwise_nocut2752_5k, seed 456
Pointwise scorer (N5): one prompt per (paragraph, candidate) pair, answer Yes/No; score = P(Yes) = softmax([Yes, No]). Trained on clean Train-5k (5,000 paragraphs), 3 random negatives per paragraph (by seed); MAX_LEN 2752 (no truncation); LR 2e-4, batch 2 x accum 8, 2 epochs, best epoch by validation F1.
- Role: Main pointwise component of CiteDuoWise (thesis).
- Base model:
google/gemma-4-E4B-it; LoRA r=16, alpha=32, dropout=0.05, target modules ['k_proj', 'o_proj', 'q_proj', 'v_proj']. - Seed: 456. Code, prompts, data split and evaluation: https://github.com/tathiyennhi/automatic-citation-checking (package
citeduowise/). - Local experiment folder:
experiments/gemma4_nocut_5k_clean/(protocol and logs).
Files (sha256)
adapter_config.json:8b4a2637cfe71d4932bbc0f0716c78924ebd03d313d3e2f2075b2cf4403b43e5adapter_model.safetensors:e7e356f8705e5df300ef43ab88e7a882f84abb963811a0c201d17594d290e76bchat_template.jinja:0a2c8073c878ab1da004bee933a998606537bbb62016310352c7285c3f01c5b5tokenizer.json:91d9cf86906f78e197021fda88214d448e5c120665d0243a8e776e046986c6bctokenizer_config.json:3ab5c7b94dc97d65ca7064496fa69b88ff875378e1cb7ee3e43070c3a8170999