Code Evidence Judge (SmolLM2 360M)
This is a full-parameter supervised fine-tune of
HuggingFaceTB/SmolLM2-360M-Instruct for one narrow,
schema-bound developer-agent subroutine:
Decide answer-vs-continue from gathered evidence snippets.
The model is one cell from the Parameter Floors for Developer-Agent Subroutines experiment. Labels are generated by deterministic oracles over real Python repositories; no teacher model or human judge labels the data.
Intended Use
Use this checkpoint inside the repository's verified subroutine harness, which renders the task-specific prompt, parses strict JSON, permits one localized schema-feedback retry, applies deterministic guards, and falls back to rules where appropriate. This is not a general coding assistant or chat model.
Evaluation
Evaluation uses up to 250 examples from HTTPX and Jinja2, both held out entirely from training. Decoding is greedy.
| Metric | Result |
|---|---|
| Success after one schema retry | 50.0% |
| First-pass success | 50.0% |
| First-pass schema validity | 99.2% |
| Base instruct success after retry | 3.2% for the base instruct model |
| Rules-only success | 98.3% |
Experiment verdict for this subroutine: rules suffice.
Training
- Training examples: 1460
- Epochs: 3.0
- Learning rate: 2e-05
- Effective batch configuration: 8 per device x 4 gradient accumulation
- Maximum sequence length: 2048
- Seed: 0
- Final training loss: 1.127464
- Reproduction hardware: one NVIDIA A100 80GB PCIe
- Source revision:
d0fd7bf
The dataset was generated from pinned Flask, Click, and Rich repositories for training/validation. HTTPX and Jinja2 were reserved for testing.
Limitations
The checkpoint is specialized to one closed JSON schema and should not be expected to retain broad instruction-following ability. The experiment mixes two base-model families across its size sweep. Some subroutines are better served by deterministic rules; consult the verdict above before deployment.
License
Apache-2.0, following the base model. Experiment code is MIT licensed.