Skip to content

Start v0.6.4 development and record the v0.6.3 release evidence #47

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
DaoyuanLi2816 merged 1 commit into from
Aug 6, 2026
Merged

Start v0.6.4 development and record the v0.6.3 release evidence #47

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion CITATION.cff
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ title: "miniVERL: Auditable single-GPU alignment and distillation runtime"
message: "If you use miniVERL in your work, please cite it as below."
type: software
version: 0.6.3
date-released: 2026-08-05
date-released: 2026-08-06
license: Apache-2.0
repository-code: "https://github.com/DaoyuanLi2816/mini-verl"
url: "https://github.com/DaoyuanLi2816/mini-verl"
Expand Down
2 changes: 1 addition & 1 deletion PROJECT_STATE.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@ and what it printed.

Last updated: 2026-08-05.

Canonical release state: releasing `v0.6.3`.
Canonical release state: stable `v0.6.3` (`005a4549da713716e64c3ae80ff55fb131519f79`), development `0.6.4.dev0`.
Every public version claim is generated from `release-state.yaml` and gated by
`python scripts/release_state.py --check`.

Expand Down
46 changes: 23 additions & 23 deletions PYPI.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
<p align="center">
<img src="https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.6.3/docs/banner.svg" alt="miniVERL — single-GPU LLM post-training" width="880">
<img src="https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/banner.svg" alt="miniVERL — single-GPU LLM post-training" width="880">
</p>

<div align="center">
Expand All @@ -8,15 +8,15 @@
[![Build](https://github.com/DaoyuanLi2816/mini-verl/actions/workflows/build.yml/badge.svg)](https://github.com/DaoyuanLi2816/mini-verl/actions/workflows/build.yml)
[![PyPI](https://img.shields.io/pypi/v/miniverl.svg)](https://pypi.org/project/miniverl/)
[![Python](https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12%20%7C%203.13-blue)](https://www.python.org)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/LICENSE)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/DaoyuanLi2816/mini-verl/blob/main/LICENSE)

</div>

<p align="center">
<a href="https://pypi.org/project/miniverl/"><strong>PyPI</strong></a> ·
<a href="https://daoyuanli2816.github.io/mini-verl/"><strong>Stable docs</strong></a> ·
<a href="https://daoyuanli2816.github.io/mini-verl/dev/">Development docs</a> ·
<a href="https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/README.zh-CN.md">中文</a>
<a href="https://github.com/DaoyuanLi2816/mini-verl/blob/main/README.zh-CN.md">中文</a>
</p>

**miniVERL is a local, inspectable runtime for a documented subset of
Expand Down Expand Up @@ -44,15 +44,15 @@ optimization in about 50 seconds on the measured laptop CPU. For inspection,
schemas and reports without the ML stack, use `pip install miniverl`. For CUDA
training, install the matching CUDA-enabled PyTorch wheel first, then install
`miniverl[train,cuda]`; the extra does not select a CUDA PyTorch build. See the
[single-GPU guide](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/single-gpu-guide.md).
[single-GPU guide](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/single-gpu-guide.md).

## Three paths

| Path | Start with | Concrete artifact | Next |
| --- | --- | --- | --- |
| **Align** — compare SFT, DPO, KD and OPD only when the pilot evidence supports the cost | `miniverl pilot recipes/alignment_policy_conditioned_qwen.yaml` | `alignment-card.json` | [Alignment Lab](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/alignment-lab/alignment-lab-v1.md) |
| **Distill locally** — strict OPD, shared backbones and padded trajectory updates on one CUDA GPU | `miniverl train recipes/qwen_consumer_gpu_shared.yaml --dry-run` | `config.resolved.yaml` plus a revision-pinned PEFT adapter | [Bring your own GPU](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/single-gpu-guide.md) |
| **Scale out** — import a documented profile, convert Parquet, export a bundle and run bridge checks | `miniverl bridge doctor scaleout-bundle` | `provenance/compatibility-report.json` | [Verified verl artifact bridge](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/verl-bridge.md) |
| **Align** — compare SFT, DPO, KD and OPD only when the pilot evidence supports the cost | `miniverl pilot recipes/alignment_policy_conditioned_qwen.yaml` | `alignment-card.json` | [Alignment Lab](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/alignment-lab/alignment-lab-v1.md) |
| **Distill locally** — strict OPD, shared backbones and padded trajectory updates on one CUDA GPU | `miniverl train recipes/qwen_consumer_gpu_shared.yaml --dry-run` | `config.resolved.yaml` plus a revision-pinned PEFT adapter | [Bring your own GPU](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/single-gpu-guide.md) |
| **Scale out** — import a documented profile, convert Parquet, export a bundle and run bridge checks | `miniverl bridge doctor scaleout-bundle` | `provenance/compatibility-report.json` | [Verified verl artifact bridge](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/verl-bridge.md) |

The bridge is a **verified artifact bridge**: a pinned config/data/model
parse-load smoke at miniVERL-defined compatibility Level 3. It has never run a
Expand Down Expand Up @@ -94,12 +94,12 @@ improved it; continued SFT and both OPD variants retained measured regressions.
| standard OPD | 98.6% | 97.2% | 100.0% | 76.7 s |
| verifier-gated OPD | 97.9% | 95.8% | 46.8% | 66.0 s |

![Alignment and utility deltas from the saturated SFT checkpoint; small marks are all three seeds and large marks are means](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.6.3/docs/alignment-lab/delta-from-sft.svg)
![Alignment and utility deltas from the saturated SFT checkpoint; small marks are all three seeds and large marks are means](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/alignment-lab/delta-from-sft.svg)

The two sandbox safety checks tied at zero while utility still regressed.
IFEval, XSTest, HarmBench and RewardBench were **not executed**. “Preference
win rate” is a deterministic Minipolicy paired outcome, not human preference.
Read the [study, seed-level values and limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/alignment-lab/alignment-lab-v1.md).
Read the [study, seed-level values and limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/alignment-lab/alignment-lab-v1.md).

## One measured systems result

Expand All @@ -110,13 +110,13 @@ reserved memory versus 3.035 GiB for dual model, while running 10.1% slower.
All 12 preregistered equivalence comparisons passed. These are one-workload,
one-machine measurements, not promises for other GPUs.

![Measured throughput and reserved VRAM for dual-model and shared-backbone runtime cells](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.6.3/docs/consumer-runtime-v1-pareto.svg)
![Measured throughput and reserved VRAM for dual-model and shared-backbone runtime cells](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/consumer-runtime-v1-pareto.svg)

[Consumer Runtime v1 methods and caveats](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/consumer-runtime-v1.md)
[Consumer Runtime v1 methods and caveats](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/consumer-runtime-v1.md)

## Compatibility boundary

![Verified local runtime, portable artifact bundle and pinned upstream smoke; distributed verl execution remains untested](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/v0.6.3/docs/verl-bridge-architecture.svg)
![Verified local runtime, portable artifact bundle and pinned upstream smoke; distributed verl execution remains untested](https://raw.githubusercontent.com/DaoyuanLi2816/mini-verl/main/docs/verl-bridge-architecture.svg)

The bridge targets official verl `v0.8.0` at commit `7aed6b23` and uses the
term **miniVERL-defined compatibility Level 3**. That means a checksummed
Expand All @@ -133,21 +133,21 @@ PPO/reward scaffold, not an executable continuation of miniVERL OPD semantics.

## Detailed studies and preserved negative evidence

- [RecoveryBench v1](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/recoverybench/recoverybench-v1.md): frozen-student KD
- [RecoveryBench v1](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/recoverybench/recoverybench-v1.md): frozen-student KD
outperformed much slower fresh-state OPD on the preregistered primary view;
the verifier gate remained `insufficient_evidence`.
- [Alignment Lab v1](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/alignment-lab/alignment-lab-v1.md): the starting SFT
- [Alignment Lab v1](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/alignment-lab/alignment-lab-v1.md): the starting SFT
checkpoint was at the ceiling, so no positive OPD result is claimed.
- [Calculator benchmark](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/benchmarking.md): both negative controls completed
- [Calculator benchmark](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/benchmarking.md): both negative controls completed
normally and measured 0% strict success. They were not configuration
failures. Because they used the historical ambiguous protocol-v1 prompt,
their failure cannot be attributed solely to intrinsic teacher behavior.
- [Consumer Runtime v1](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/consumer-runtime-v1.md): padded update batches and
- [Consumer Runtime v1](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/consumer-runtime-v1.md): padded update batches and
shared adapters preserve the measured one-update objective within declared
tolerances; rollout generation remains sequential.
- [Limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/limitations.md), [math](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/math.md),
[reproducibility](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/reproducibility.md) and
[compatibility policy](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/docs/compatibility.md).
- [Limitations](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/limitations.md), [math](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/math.md),
[reproducibility](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/reproducibility.md) and
[compatibility policy](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/compatibility.md).

New runs establish tokenizer compatibility through structural identity. The
legacy behavioral fingerprint—token IDs for one fixed probe plus metadata—is
Expand All @@ -167,7 +167,7 @@ python -m pip install -e ".[dev]"
pytest -q -m "not gpu and not network"
```

Apache-2.0 licensed. See [CONTRIBUTING.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/CONTRIBUTING.md) and
[SECURITY.md](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/SECURITY.md). Project records: [default GPU recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/recipes/qwen_consumer_gpu_calc.yaml),
[frozen calculator JSON](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/benchmarks/results/gpu-calc-hard-equal-update-v2.json),
[changelog](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/CHANGELOG.md), [citation](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/CITATION.cff) and [license](https://github.com/DaoyuanLi2816/mini-verl/blob/v0.6.3/LICENSE).
Apache-2.0 licensed. See [CONTRIBUTING.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CONTRIBUTING.md) and
[SECURITY.md](https://github.com/DaoyuanLi2816/mini-verl/blob/main/SECURITY.md). Project records: [default GPU recipe](https://github.com/DaoyuanLi2816/mini-verl/blob/main/recipes/qwen_consumer_gpu_calc.yaml),
[frozen calculator JSON](https://github.com/DaoyuanLi2816/mini-verl/blob/main/benchmarks/results/gpu-calc-hard-equal-update-v2.json),
[changelog](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CHANGELOG.md), [citation](https://github.com/DaoyuanLi2816/mini-verl/blob/main/CITATION.cff) and [license](https://github.com/DaoyuanLi2816/mini-verl/blob/main/LICENSE).
2 changes: 1 addition & 1 deletion docs/generated/quality.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
{
"schema_version": 1,
"release": "0.6.3",
"status": "validated",
"status": "released",
"measured_commit": "e8a36db60b2146b684204df8a8bea3e20c11d8e4",
"measured_at": "2026-08-05T18:20:00-07:00",
"platform": "Windows 11 Pro 10.0.22631, CPython 3.12, coverage branch mode",
Expand Down
4 changes: 2 additions & 2 deletions docs/overrides/main.html
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,12 +1,12 @@
{% extends "base.html" %}

{% block announce %}
<div class="docs-channel" data-stable-version="0.6.3" data-dev-version="0.6.3">
<div class="docs-channel" data-stable-version="0.6.3" data-dev-version="0.6.4.dev0">
<strong id="docs-channel-label">Stable documentation</strong>
<label for="docs-version-selector">Version</label>
<select id="docs-version-selector" aria-label="Documentation version">
<option value="stable">Stable 0.6.3</option>
<option value="dev">Development 0.6.3</option>
<option value="dev">Development 0.6.4.dev0</option>
</select>
</div>
{% endblock %}
26 changes: 26 additions & 0 deletions docs/release-checklist.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,14 @@ This is the release gate and publication record for miniVERL. A checked item
names an invariant exercised on the stated source. Publication begins only
after the exact release commit and its remote checks are green.

## v0.6.4 (in development)

- [ ] A reward scaffold saved with a UTF-8 byte-order mark parses instead of
being reported as a syntax error. CPython strips the BOM when it reads a
source file; the static checker did not, so a scaffold written by a
Windows editor was refused as unparseable. Found while verifying the
published v0.6.3 wheel.

## v0.6.3 Security, artifact integrity and release-state hardening

- [x] `bridge doctor` executes zero Python from the inspected bundle by
Expand Down Expand Up @@ -38,6 +46,24 @@ after the exact release commit and its remote checks are green.
crash atomicity is claimed.
- [x] Every frozen scientific artifact remains byte-identical.

### After the v0.6.3 tag

- [x] Annotated tag `v0.6.3` resolves to release commit
`005a4549da713716e64c3ae80ff55fb131519f79`; release run
[`31084165317`](https://github.com/DaoyuanLi2816/mini-verl/actions/runs/31084165317)
completed OIDC Trusted Publishing and GitHub Release creation.
- [x] PyPI and the GitHub Release expose identical files. Wheel SHA-256
`1c620c310e8e4156f515d52128a0f26a037096347a2bc105c63f173e6563f5e0`,
sdist SHA-256
`b93d2fad63432ecd680b72fe0bf79f6b0be212af1aab82780ed8e31636edeefe`.
- [x] The PyPI integrity API exposes one Trusted Publisher attestation bundle
per distribution, bound to `DaoyuanLi2816/mini-verl` and `release.yml`.
- [x] A clean Windows Python 3.12 install from `https://pypi.org/simple`
reported `miniverl 0.6.3`, kept torch absent, and refused to execute a
hostile reward scaffold: the top-level marker write never ran and the
doctor reported `top_level_call` with `code_executed: false`.
- [x] This separate state-sync change advances development to `0.6.4.dev0`.

## v0.6.1 Visual integrity and bridge correctness release

- [x] Alignment Lab publication is generated as one diverging forest chart and
Expand Down
10 changes: 4 additions & 6 deletions release-state.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,15 +17,13 @@
# such distinction, which is how its tag shipped a docs selector still
# advertising "Stable 0.6.1 / Development 0.6.2.dev0".
schema_version: 1
phase: release
phase: development

stable:
version: "0.6.3"
tag: "v0.6.3"
# This tree is the release, so it cannot name its own merge commit yet. The
# post-release state-sync replaces this with the exact tagged SHA.
release_commit: "pending"
released_at: "2026-08-05"
release_commit: "005a4549da713716e64c3ae80ff55fb131519f79"
released_at: "2026-08-06"

development:
version: "0.6.3"
version: "0.6.4.dev0"
2 changes: 1 addition & 1 deletion src/miniverl/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,6 @@

from __future__ import annotations

__version__ = "0.6.3"
__version__ = "0.6.4.dev0"

__all__ = ["__version__"]