Debug why your model learned the wrong thing.
A local GRPO evidence workspace: timeline → completion → reward patch → reviewed regression test.
简体中文 · Download v0.2.0 · Installation options · Integrations
Your reward curve is going up. Is your model actually doing the task better, or just getting better at satisfying the scorer?
PTM lets you follow a suspicious metric back to saved completions, inspect reward components, edit a declarative reward rule, and recompute ranks and GRPO advantages on the same answers. Turn the expected ordering into a reviewed contract and keep the incident as a reproducible offline report.
Requires 64-bit Python 3.11 or 3.12. Download the wheel from Releases. In the download directory:
macOS / Linux:
python3 -m venv .venv
.venv/bin/python -m pip install ./posttrain_timemachine-0.2.0-py3-none-any.whl
.venv/bin/ptm demoWindows PowerShell:
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install .\posttrain_timemachine-0.2.0-py3-none-any.whl
.\.venv\Scripts\ptm.exe demoOpen http://127.0.0.1:8765. No activation script, Node.js, API key, account or pretrained-model download is needed for the demo. Initial dependency installation needs the internet; the running app uses bundled local assets. Stop with Ctrl+C. Data lives under .ptm/ in your working directory.
Other options: download the source ZIP and install with pip install ., clone the repository, use pip install "git+https://github.com/renweimeng/posttrain-timemachine.git@v0.2.0" inside a venv, or run uvx --python 3.12 --from ./posttrain_timemachine-0.2.0-py3-none-any.whl ptm demo after downloading the wheel. There is no PyPI release or native desktop installer yet.
| Environment | Application | Deep training / update replay |
|---|---|---|
| macOS | Local end-to-end and isolated-wheel verification; Python 3.11/3.12 CI | Actual CPU reference training and two-process capture tested |
| Linux | Python 3.11/3.12 installation and core/browser CI | Separate pinned TRL reference job passed; GPU requires its own validation |
| Windows | Native PowerShell installation smoke passed on Python 3.11/3.12 | Use WSL2/Linux for the reference examples; native deep training is unverified |
The current release provides a Python wheel and source archives, not an .exe or .dmg installer. Prebuilt frontend assets are included, so ordinary users do not need Node.js. Mobile layout support is for the browser interface, not a phone-hosted Python server.
- Select demo-format-truth.
- Click Reveal Evidence on the timeline.
- Inspect the wrong-but-perfectly-formatted answer and its reward breakdown.
- Open What-if Lab and reduce Format weight from
0.7to0.05. - Recompute Deltas and inspect rank reversals and advantage sign flips.
- Review the expected fixture ordering and Save reviewed contract.
- Run it in Regression Tests and export a report from Reports.
Four complete authored Arena challenges cover format over truth, zero signal, verbosity and checker blind spots. Synthetic examples are labeled; real CPU trainer traces are separate.
The update view uses an actual tiny random-model CPU capture. It checks one real optimizer step, not a future model-quality prediction.
Import canonical JSON/JSONL, explicit-field verl JSONL or complete rank bundles in Runs & Import. Try the actual tiny CPU trace. Browser imports are limited to 4 MiB; larger files use ptm import trace.jsonl --format ptm-jsonl (128 MiB per file, 2 MiB per JSONL line).
Capture explicit run/batch/prompt-group IDs, completions, rewards and the inputs required by your reward rule. Plain curves or arbitrary trainer logs do not contain enough evidence. After import, use Load recorded reward before editing a rule. See integrations for actual trainer capture, two-rank CPU capture and retention options.
| Saved evidence | Available operation |
|---|---|
| Complete groups, rewards and scoring inputs | Reward / rank / advantage diff |
| Token IDs, masks and staged policy logprobs | Token signals and supported full-batch scalar objective replay |
| Separate supported numeric checkpoint capture | One real CPU optimizer step from fixed inputs |
- Actual reference: TRL 0.24.0 with source-fingerprinted opt-in deep capture.
- Loss references: GRPO, BNPO, DAPO and Dr.GRPO; full rank/microbatch denominators have independent tests.
- Native deep capture: CPU, one process, accumulation=1, iterations=1 and supported loss settings.
- Update replay: bounded local GPT-2, safetensors, native AdamW/SGD; no arbitrary checkpoint upload.
- A real two-process CPU L1 integration is included; native distributed L2 is not validated.
- SFT/DPO/PPO objectives, general Agent framework hooks and future training quality are not automatically supported.
Local acceptance on 2026-11-03: 81 Python tests, 12 browser journeys, 84.78% statement coverage, real CPU training/updates, offline L2 recomputation and an isolated wheel install without Torch. macOS was verified locally; Linux and Windows smoke coverage is defined in CI, with actual results visible in Actions. Native Windows deep training and GPU throughput remain unverified.
Read test evidence, numerical correctness, raw benchmarks and security scope. The pinned optional reference extra has known dependency advisories and is separate from the default application.
The server is loopback-only and has no login, telemetry or multi-tenant hosting mode. Text retention is configurable; retaining numeric token IDs is not anonymization. Ordinary report bundles exclude weights and optimizer state.
Bring a minimal reproducible reward failure, reviewed contract, versioned adapter or real reference fixture. See CONTRIBUTING. If this helps you catch a training bug, a Star or a reproducible issue is welcome.
Apache-2.0. Sample provenance · Architecture · Feature boundaries.
Do I need an account? No. Run the application yourself and open the local browser URL. The current app is single-user and loopback-only.
Where are my records? In .ptm/ under your working directory, unless you choose another root. Start ptm serve .ptm --open from the same directory to reopen it. Keep the terminal running; Ctrl+C stops the server.
The port is busy. Use ptm demo --port 8767 or ptm serve .ptm --port 8767, then visit port 8767.
The command is missing. Invoke .venv/bin/ptm on macOS/Linux or .\.venv\Scripts\ptm.exe on Windows, as shown above. Activation is optional.
Can I inspect a remote training run? Download its captured trace, or start PTM on the server and forward the loopback port with ssh -L 8765:127.0.0.1:8765 user@server.example.com. Then open localhost:8765 on your computer. Replace the account and hostname with your own.
Can I import a dataset or a screenshot of a reward curve? Full analysis needs recorded completions, rewards and explicit group identity. A dataset or curve alone does not reconstruct missing training evidence.
How do I verify a download? The release includes SHA256SUMS for the named wheel and uploaded source ZIP. Use shasum -a 256 FILE on macOS, sha256sum FILE on Linux, or Get-FileHash FILE -Algorithm SHA256 in PowerShell and compare the value. GitHub's automatically generated source archives are separate downloads.



