Gemma-4 E4B LoRA — SCIDOCA 2025 Task 1 (citation discovery) — 47k, seed 123
Pointwise scorer trained on the FULL clean Train (47,166 paragraphs), MAX_LEN 512, up to 3 negatives per paragraph; LR 2e-4, effective batch 16, 2 epochs fixed, warm-up 10%, checkpoint 'best' by validation pair F1.
- Role: Supplementary run, NOT used in the thesis results; kept for possible follow-up experiments.
- 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: 123. Code, prompts, data split and evaluation: https://github.com/tathiyennhi/automatic-citation-checking (package
citeduowise/). - Local experiment folder:
experiments/gemma4_fulldata_clean/(protocol and logs).
Files (sha256)
adapter_config.json:c4dc5a8b2f91b1da26230c96a948e3db1e79b4d8318d5f77335c0a28aeb9a454adapter_model.safetensors:fe41ac0a1d487b1c5a4f3baaeb687423a450c38117a4d23987feee0d5c7e15bdchat_template.jinja:0a2c8073c878ab1da004bee933a998606537bbb62016310352c7285c3f01c5b5tokenizer.json:91d9cf86906f78e197021fda88214d448e5c120665d0243a8e776e046986c6bctokenizer_config.json:3ab5c7b94dc97d65ca7064496fa69b88ff875378e1cb7ee3e43070c3a8170999