8000 Release v0.2.6 lifecycle and privacy correctness by DaoyuanLi2816 · Pull Request #25 · DaoyuanLi2816/mini-verl · GitHub
[go: up one dir, main page]

Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 27 additions & 1 deletion CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,31 @@ All notable changes to miniVERL are recorded here. The format follows

## [Unreleased]

## [0.2.6] - 2026-08-01

Small concurrency, lifecycle and privacy correctness release. No training
objective, environment, benchmark, model family, adapter revision or frozen
scientific result changed.

### Changed

- Training, evaluation, checkpoint save/load and destructive close now share
one non-blocking trainer-operation ownership contract. Checkpoint loading is
READY-only both before and after ownership acquisition.
- Evaluation records the actual prior model mode and restores it in `finally`
after success or any rollout, serialization, diagnostics, metrics or event
failure.

### Fixed

- `close()` can no longer release or null resources while evaluation,
checkpoint save/load or training owns them; failed ownership attempts mutate
no state or artifact, while later and repeated close remain safe.
- Portable artifacts recognize credential semantics across common key styles,
structurally redact userinfo in HTTP/SSH/database URLs, and sanitize embedded
absolute Windows, UNC and POSIX paths without rewriting public URLs,
relative paths or mathematical slash expressions.

## [0.2.5] - 2026-07-30

Focused correctness release. No training objective, environment, benchmark,
Expand Down Expand Up @@ -401,7 +426,8 @@ Same-tokenizer only; one trajectory per forward pass; `swap` unavailable for
quantized models; only Qwen3 and Qwen2 architectures tested; single-seed GPU
results. The full list is in `docs/limitations.md`.

[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.5...HEAD
[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.6...HEAD
[0.2.6]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.5...v0.2.6
[0.2.5]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.4...v0.2.5
[0.2.4]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.3...v0.2.4
[0.2.3]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.2.2...v0.2.3
Expand Down
6 changes: 3 additions & 3 deletions CITATION.cff
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,8 @@ cff-version: 1.2.0
title: "miniVERL: On-policy distillation for tool-using agents on one GPU"
message: "If you use miniVERL in your work, please cite it as below."
type: software
version: 0.2.5
date-released: 2026-07-30
version: 0.2.6
date-released: 2026-08-01
license: Apache-2.0
repository-code: "https://github.com/DaoyuanLi2816/mini-verl"
url: "https://github.com/DaoyuanLi2816/mini-verl"
Expand Down Expand Up @@ -98,4 +98,4 @@ references:
url: "https://arxiv.org/abs/2603.07079"
notes: >-
Motivates recording per-token teacher entropy. Entropy-aware divergence
mixing is a roadmap item and is not implemented in miniVERL v0.2.5.
mixing is a roadmap item and is not implemented in miniVERL v0.2.6.
19 changes: 18 additions & 1 deletion PROJECT_STATE.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,24 @@ Living build log for **miniVERL** (`mini-verl` / `miniverl` / CLI `miniverl`).
A checkbox is not evidence: every completed item names the command that was run
and what it printed.

Last updated: 2026-07-31.
Last updated: 2026-08-01.

## v0.2.6 concurrency, lifecycle and privacy correctness release

| item | current state |
| --- | --- |
| audited baseline | fetched all remotes and started `v0.2.6-final-concurrency` from clean public `main` at `ba0cbc33d731936dcb0e44fc42e29cae2f1c803d` (`0.2.6.dev0`) |
| regression-first evidence | isolated pre-fix runs reproduced a real `close()` teardown race, six checkpoint-load overlap failures, six evaluation-mode restoration failures, and 22 semantic-key/URL/path privacy failures; deterministic concurrency tests use `threading.Event`, never timing sleeps |
| ownership and mode | `close()` owns the same non-blocking operation guard as train/evaluate/checkpoint work before changing state or releasing resources; public checkpoint load checks READY before and after ownership and delegates to an owner-only implementation; evaluation restores the exact prior model mode in `finally` |
| privacy | semantic components cover snake/kebab/camel/Pascal keys with explicit benign metadata exceptions; supported URL schemes retain public structure while removing userinfo; arbitrary embedded absolute Windows/UNC/POSIX paths are reduced to portable basenames without rewriting URLs, relative paths or mathematical slash expressions |
| focused adversarial pass | lifecycle, checkpoint-resume, standalone-evaluation, benchmark privacy and all shareable report-format tests pass **209 tests**, including generated semantic keys, credentialed URLs and POSIX paths; an added false-positive probe found and fixed mid-field matching in `loss/token = 0.5` |
| static and documentation gates | `git diff --check`, Ruff check/format, mypy over 77 source files, actionlint 1.7.12 with verified upstream checksum, tag-pinned `PYPI.md`, Markdown links, generated schema and generated benchmark-SVG byte comparison all pass |
| complete local suite | after refreshing editable distribution metadata from `0.2.6.dev0` to source `0.2.6`, the full non-GPU/non-network gate passed **1348 tests**, 6 deselected, at **87.46%** branch coverage; the available RTX 4080 gate passed **5** and the network gate passed **3** |
| compatibility evidence | torch-free core passed **1058 tests** on each of Python 3.10, 3.11, 3.12 and 3.13; the minimum bundle (Torch 2.3.1, Transformers 4.51.3, PEFT 0.12.0, Accelerate 0.33.0, NumPy 1.24.4, bitsandbytes 0.43.3) and latest bundle (Torch 2.13.0, Transformers 5.14.1, PEFT 0.20.0, Accelerate 1.14.0, NumPy 2.5.1, bitsandbytes 0.50.0) each passed 2 no-network training smokes and the **133-test** HF/config suite |
| platform concurrency | the Windows complete suite includes the deterministic operation matrix; a fresh WSL2/Linux Python 3.12 environment at implementation commit `7410674802fa178fdde1ed5064c17608cd3fa806` passed **76** lifecycle, standalone-evaluation and resume/concurrency tests; final Ubuntu PR CI remains required |
| packaging and installs | one `0.2.6` wheel and one sdist pass Twine; the extracted sdist passes Ruff, format, mypy and **1348** CPU tests, rebuilds a wheel with the same **78-file** runtime inventory, and clean Python 3.10 core plus Python 3.12 `[train]` installs pass version/doctor and demo/inspect/report/standalone-eval smokes |
| immutable evidence | the frozen calculator JSON remains byte-identical at SHA-256 `53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc`; all existing tags and `DaoyuanLi/mini-verl-qwen3-1.7b-protocol-teacher@23323751318135484c06c043b1f9b9e7016dd89f` remain unchanged |
| release status | metadata now identifies the validated candidate as `0.2.6`; final gates, PR, merge, tag publication and public verification remain pending |

## v0.2.5 final correctness release status

Expand Down
Loading
0