Skip to content

v0.2.0 · Debug the reward. Understand the signal.

Latest

Choose a tag to compare

@renweimeng renweimeng released this 12 Nov 10:25

PostTrain Time Machine v0.2.0

Reward 涨了,模型真的学对了吗?PTM 把“看曲线”接到“查回答、改奖励、留测试”的完整流程。

Download / 下载

  • Wheel: posttrain_timemachine-0.2.0-py3-none-any.whl — bundled local web UI; ordinary users need no Node.js or GPU.
  • Source ZIP: posttrain-timemachine-0.2.0-source.zip — source, examples, fixtures, documentation and prebuilt frontend.
  • SHA256SUMS — SHA-256 checksums for both files.

Use 64-bit Python 3.11 or 3.12. See README for macOS, Linux and Windows PowerShell commands. No PyPI upload or native desktop installer is included.

Included

  • Timeline → completion inspection → declarative reward patch → rank/advantage diff.
  • Four complete synthetic Reward Arena challenges, reviewed contracts, pytest/JUnit export and offline reports.
  • L2 Token Signal Lens, mask-aware ratio/KL/surrogate inspection and complete-batch objective replay.
  • Actual bounded CPU GPT-2 AdamW/SGD replay from numeric state, compared with recorded native after-step weights.
  • Explicit-ID TRL reference capture, actual two-rank CPU L1 fixture, explicit-field verl and recorded multi-turn import.
  • Persistent drafts, evidence links, analysis/report history, responsive layout and accessibility checks.

Verification and scope

2026-11-12 local release check: 82 Python tests and 12 browser journeys passed, along with Ruff, mypy, ESLint, TypeScript and frontend build. The wheel is verified in an isolated environment without Torch. Cross-platform install smoke workflows cover Linux, macOS and Windows; their actual execution status is shown in GitHub Actions.

Native reference uses pinned TRL 0.24.0. GPU execution, generic trainer compatibility, native Windows deep capture and long-term model improvements are not claimed. The optional pinned training extra has known dependency advisories; see docs/SECURITY.md. Ordinary application dependencies are separate.

This is a local, single-user tool without login or multi-tenant cloud hosting. All synthetic examples are labeled. Fixed-answer reward edits and one-step replay are evidence for investigation, not predictions of future model quality.

Release fixes: Windows durability flush uses a writable descriptor; the Token scroll region supports keyboard focus; cloud FPS is retained as a measurement, with an explicit controlled-hardware performance gate.

Verified GitHub Actions: all 4 core matrix jobs and the TRL reference job passed; all 6 installation smoke jobs passed on Windows, macOS and Linux with Python 3.11/3.12.