LoRA adapter specializing Qwen/Qwen3.5-2B for ATLAS document-filing JSON
(type, defining date + role, parties, evidence spans, description, needs_review).
Stacked on frozen btbtyler09/shrew-2b doc_processing,
then ATLAS LoRA r=16 / α=32 (120 CPU steps, 600 synthetic chat examples).
Usage: load shrew doc_processing first, then this adapter as atlas.
| Metric | Value |
|---|---|
| train_loss | 0.184 |
| loss step 5 → 120 | 1.875 → 0.049 |
| smoke json_ok (n=8 synthetic) | 1.0 |
| smoke field_micro (n=8) | 1.0 |
Not InternBench / realdocs screened. Do not deploy to production.
adapter_model.safetensors (~42 MB) is at repo root (sha256 b73c81620ed82ebe48eca00010318f2c94f9c5f619506c586a9e428da89e1377).
Optional decoded-binary chunk backups live under weights/chunks/ (part000.b64–part083.b64; do not base64-decode again when concatenating).
Qwen/Qwen3.5-2B