File size: 1,886 Bytes
adc02fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
# 03 Code Fixes

Date: 2026-06-19 UTC

## Installation Fixes

- Moved `pyarrow` from required dependencies to optional `parquet` extra in `pyproject.toml`.
- Reason: this cluster resolves `pyarrow` to a dummy wheel that requires an external Arrow module and breaks core `pip install -e .`.
- Core JSONL dataset path remains fully supported. Parquet remains available with `pip install -e ".[parquet]"` in an environment with Arrow.

## New Reproducible Shell Commands

Added:
- `scripts/smoke_test.sh`
- `scripts/run_train_debug.sh`
- `scripts/run_inference.sh`
- `scripts/run_eval.sh`

These scripts run tiny toy-backend jobs with `OPENCLAUDE_MOCK=1` by default.

## New Inference Command

Added:
- `scripts/infer_toy_policy.py`

It loads a CIL dataset and checkpoint, runs model policy inference when torch/model weights are available, and otherwise emits a clearly labeled fallback action selection. Model-policy toy outputs are bound to actual group targets for readable/actionable JSON.

## Evaluation Fixes

- Fixed CausalStress policy rollout crashes caused by decoded actions containing `predicted_target`.
- Added task-aware binding before toy simulator execution.

## Runtime Stability Fixes

- Added `DOVLA_TORCH_THREADS` handling in trainer/eval with default `1` for stable CPU smoke runs on shared clusters.
- Full test suite dropped from a long/hanging run to `126 passed, 1 skipped in 16.21s`.

## README Updates

- Added the four paper-audit shell commands to README quickstart.
- Documented that no-torch fallback paths are explicit and should not be interpreted as learned model results.

## Verification

- `outputs/audit_venv/bin/python -m pip install -e .` succeeded.
- `DOVLA_TORCH_THREADS=1 outputs/audit_venv/bin/python -m pytest -q` passed: `126 passed, 1 skipped`.
- Smoke, debug train, inference, eval, and expert-only baseline scripts all ran successfully.