fresh-doc (M2, prod)

Fine-tune of microsoft/unixcoder-base (125M) for FreshDoc docstring-staleness detection (Dual-Head Cross-Encoder, slot 4 full).

  • Input: (sig_old, diff_code, doc_old), max_len 1024, doc_budget 384.
  • Output: p_stale + drift_type (8 classes). Threshold tau = 0.8743, temperature T = 2.3847.
  • Train: real + AST + LLM (P1/P2 w=0.5), 3 epochs, seed 42.
  • Metrics: val PR-AUC 0.9724, test_real 0.9837, gold 0.478.

Files: model.safetensors, config.json, dualhead_config.json, tokenizer.json, tokenizer_config.json.

Use: load as AutoModel encoder + linear heads (see FreshDoc src/models/m2_model.py).

Downloads last month
22
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for o1uh/fresh-doc

Finetuned
(18)
this model