eulogik/TinyDoc-VLM-LoRA

---

TinyDoc-VLM LoRA Checkpoint

⚠️ RETIRED — adapter makes outputs worse

This LoRA adapter is part of the failed training line of the 256M checkpoint. In paired evaluation it degrades generation into literal degeneration ("$ "$ "$", 22222) compared even already-broken base weights, which themselves measure 0.0% OCRBench (n=1,000). Loss values below are training loss only — they never correlated with output quality. Kept for reproducibility; do not use for extraction.

What actually ships

The working product is the local grounded-extraction SDK running on free ollama:qwen2.5vl:3b — measured SROIE field F1 0.870 vs 0.376 for the free PP-OCR+heuristics competitor (same 100 docs, same scorer): github.com/eulogik/TinyDoc-VLM.

Historical training record (2026-06)

Parameter	Value
Base model	eulogik/TinyDoc-VLM-256M (retired)
LoRA rank / alpha	16 / 32
Trainable params	2,727,936 (0.93%)
Target modules	q_proj, v_proj, k_proj, o_proj
Data	3,000 synthetic documents (6,815 QA pairs)
Steps / best step	17,000 / 14,000 (train loss 15.0)
Hardware / time	Apple M4, 15.1 h

Training loss trajectory (loss only — not a quality metric): 43.3 → 25.7 → 20.9 → 18.6 → 16.5 → 15.0 (best, step 14k) → 17.2 (final)

License

Apache 2.0. Same as base model.

---

Part of the TinyDoc-VLM project.