miniVERL is licensed under Apache-2.0 (see LICENSE).
None. Every source file under src/miniverl/ was written for this project.
No source code from another repository has been copied, vendored or adapted.
This matters for two projects in particular, because miniVERL is positioned next to them:
- verl (
https://github.com/volcengine/verl, Apache-2.0). No verl runtime code, documentation wording, logo or visual identity is used here. The reduced test fixturetests/fixtures/verl/opd-v0.8-official-example-fields.yamlis derived from the field surface of the pinned v0.8.0 Qwen3 OPD example and retains source, commit and license provenance. miniVERL is an independent implementation of a much narrower scope and is not endorsed by verl. - OPSD (
https://github.com/HJSang/OPSD_OnPolicyDistillation). This repository has no LICENSE file and is therefore all-rights-reserved; nothing from it has been copied. It is cited indocs/comparisons.mdanddocs/references.mdas prior art only.
Implementing a published formula from its mathematical description is not
copying code, but the sources deserve credit. Full citations are in
docs/references.md.
| What | Source |
|---|---|
| The temperature-squared gradient correction on the distillation term | Hinton, Vinyals and Dean, Distilling the Knowledge in a Neural Network, arXiv:1503.02531 |
| Generalized Jensen-Shannon divergence and training on student-sampled sequences | Agarwal et al., GKD, arXiv:2306.13649 |
Teaching a student on its own trajectories against a context-conditioned teacher (the privileged_context mode) |
Ye, Dong, Wu, Huang and Wei, On-Policy Context Distillation for Language Models, arXiv:2602.12275 |
| Recording per-token teacher entropy as a first-class metric | Jin et al., Entropy-Aware On-Policy Distillation of Language Models, arXiv:2603.07079 |
Numerically stable log(1 - exp(x)) in two regimes |
Machler, Accurately Computing log(1 - exp(-|a|)), R Rmpfr vignette, 2012 |
| NF4 quantization with double quantization, and LoRA adapters on a quantized base | Dettmers et al., QLoRA, arXiv:2305.14314; Hu et al., LoRA, arXiv:2106.09685 |
The top-k plus tail coarse-graining of a teacher distribution is not novel
either; TRL's ServerDistillationTrainer exposes loss_top_k with an optional
tail bucket. miniVERL's contribution there is naming it accurately and proving
the lower-bound relationship in tests, not inventing it.
Installed from PyPI, not redistributed in this repository. Each remains under its own license; consult the installed distribution for the authoritative text.
| Package | License | Used for |
|---|---|---|
| typer | MIT | CLI |
| rich | MIT | terminal rendering |
| pydantic | MIT | configuration and schema validation |
| PyYAML | MIT | recipe parsing |
| Jinja2 | BSD-3-Clause | HTML report template |
| platformdirs | MIT | platform-correct paths |
| safetensors | Apache-2.0 | the pickle-free tensor format used by the cache and checkpoints |
torch (extra train) |
BSD-3-Clause | tensors, autograd |
transformers (extra train) |
Apache-2.0 | causal-LM backend |
peft (extra train) |
Apache-2.0 | LoRA / QLoRA adapters |
accelerate (extra train) |
Apache-2.0 | device placement helpers |
numpy (extra train) |
BSD-3-Clause | array interop |
bitsandbytes (extra cuda) |
MIT | NF4 quantization, 8-bit optimizer |
pytest, hypothesis, jsonschema, ruff, mypy, build, twine (extra dev) |
MIT / Apache-2.0 / BSD | development only, not redistributed |
miniVERL downloads no weights by default. The published recipes name two models, both Apache-2.0, pinned by revision:
Qwen/Qwen3-0.6Batc1899de289a04d12100db370d81485cdf75e47caQwen/Qwen3-1.7Bat70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
No model weights are committed to this repository, and .gitignore refuses the
common weight extensions so they cannot be added by accident.
The v0.7.0 external alignment study evaluates against four upstream datasets and two upstream evaluator models. None of their content is redistributed here. The repository commits identifiers, pinned revisions, selected row ids, content digests and resulting scores; prompts and model generations stay upstream, and harmful-compliance generations are stored only as SHA-256 digests.
| Asset | Licence | Pinned revision |
|---|---|---|
google/IFEval |
Apache-2.0 | 966cd89545d6b6acfd7638bc708b98261ca58e84 |
natolambert/xstest-v2-copy |
CC-BY-4.0 | b71afe2a6d10e5a6254ea8bcb006c48b095a15d5 |
JailbreakBench/JBB-Behaviors |
MIT | 886acc352a31533ffbcf4ef22c744658688086fc |
allenai/reward-bench |
ODC-By | 168d848cdbbea9764fae4a544dc9ca1e6cca4931 |
ibm-granite/granite-guardian-3.0-2b |
Apache-2.0 | e48b7b8acf438d24daa2271ada6df945b5b8895e |
llm-blender/PairRM |
MIT | 5b880cc73776ac75a835b3e0bd5169bcb5be013b |
The IFEval scorer in src/miniverl/alignment_external/ifeval.py is an
independent implementation written from the published instruction semantics. No
official scorer is installable from PyPI and none was copied; where it
substitutes for a reference dependency the affected instruction types are
marked approximate in every result.
The XSTest refusal classifier is likewise independent, and is the paper's string-matching variant rather than its GPT-4 judge configuration. Neither endpoint is reported under a benchmark name whose official evaluator was not run — see benchmark governance.
docs/banner.svg was drawn for this project. It uses no third-party artwork,
font file, icon set or trademark. The Contributor Covenant text in
CODE_OF_CONDUCT.md is version 2.1, distribute
2D60
d under
CC BY 4.0 by the Contributor Covenant project.