Skip to content

verification.md

Latest commit

 

History

History
54 lines (37 loc) · 3.61 KB

verification.md

File metadata and controls

54 lines (37 loc) · 3.61 KB
Raw

Local verification record

Date: 2026-10-20 (Asia/Shanghai).

Environment and provenance

  • Windows; Python 3.12.14; PyTorch 2.11.0+cu128; Transformers 5.19.0.
  • NVIDIA GeForce RTX 3060; CUDA available.
  • Reused E:/Code/microduck-qwen/.venv/Scripts/python.exe; no dependencies were installed or changed.
  • Set PYTHONPATH=E:/Code/nano-posttrain, HF_HUB_OFFLINE=1, and WANDB_DISABLED=1 for CLI training.
  • Used cached Qwen/Qwen2.5-0.5B snapshot 060db6499f32faf8b98477b0a26969ef7d8b9987.
  • The default Anaconda interpreter has no torch. The reused torch environment has no pytest, so tests used the existing pure-Python pytest installation from D:/anaconda3/Lib/site-packages, appended after the environment's existing import paths.

The README training commands specify the actual data, outputs, and flags. Retained logs: SFT, DPO. Checkpoints are local under output/readme_sft/step_latest and output/readme_dpo/latest; binary model artifacts are not part of this documentation change.

Tests

Default invocation through pytest's Python API:

import sys
sys.path.append("D:/anaconda3/Lib/site-packages")
import pytest
raise SystemExit(pytest.main(["tests", "-q", "-p", "no:cacheprovider"]))

Result: 13 passed, 1 failed, 3 skipped in 2.44 seconds. Failure: test_sample_sft_json_loads expects the obsolete top-level sample-data path. No tests were weakened, skipped beyond their existing markers, or changed.

Final smoke invocation used offline cached model access:

from transformers import AutoModelForCausalLM, AutoTokenizer
import sys
sys.path.append("D:/anaconda3/Lib/site-packages")
import pytest
raise SystemExit(pytest.main([
    "tests/test_smoke.py", "-q", "--run-smoke",
    "-p", "no:cacheprovider", "--basetemp=output/pytest-readme-smoke"
]))

Result: 3 passed, 2 pytest plugin rewrite warnings in 1.94 seconds. An earlier attempt encountered a temporary-directory permission error and incompatible optional scikit-learn discovery from the Anaconda path. Preloading Transformers before adding that path and using a workspace temporary directory resolved the harness issues. No model test failure remains in the final smoke invocation.

Training interpretation

Both CLI runs completed two optimizer steps and wrote adapter checkpoints. SFT losses: 2.7812 and 1.4688, on different batches. DPO losses: 0.7027 and 0.7157, with logged preference accuracy 0.00. First-step learning rate was zero during warm-up. DPO gamma was its CLI default, zero; reference device was CPU.

These results establish execution, not generalization or training benefit. No scored generation, TRL comparison, quantization, serving, full convergence, or peak GPU-memory test was performed.

README figure and engineering review

The README figure is generated by python docs/generate_readme_charts.py using Matplotlib (rendered here with 3.7.2). It reads the retained SFT/DPO log points and parameter counts directly; test_summary.json transcribes the test results recorded above. No smoothing, interpolation, aggregation of losses, or uncertainty interval is applied. Default and smoke outcomes represent separate invocations.

The 1920 x 1520 PNG was visually inspected. Figure generation, malformed-input rejection, English documentation, relative Markdown links, code fences, and git diff --check were checked. No new GPU training was needed to render existing results. See the software engineering review for the source-based findings and proposed checks.