Avifenesh/k2h-evals

K2H evaluation sets

Evaluation data used for continued pretraining of IFM/K2-Horizon-MoVA-36B-A4B toward LLM inference and ML systems engineering. All of it comes from public upstream sources dated after the training corpus cutoff. Companion training data: Avifenesh/k2h-corpus-domain-clean.

Published 2026-11-08.

Contents

Subset	What	Records	Tokens
pr-validated	Validated PR eval (version 3): does this fix pass validation?	154 items	565,131 prompt tokens
pr-validated-nll	Text of the PRs behind the eval, for held-out NLL	193 PRs	1,306,776
pr-validated-meta, pr-validated-prs	Labels, tiers and evidence per item and per PR	154 / 193
heldout-prs	Merged PRs from 9 inference repositories, 2026-10-16 to 2026-10-30	4,179	45,035,928
heldout-commits	Commits from 24 inference and kernel repositories, 2026-10-16 to 2026-10-30	1,349	10,365,363
heldout-papers	arXiv ML-systems papers dated 2026-10-01 to 2026-10-22, CC BY 4.0, CC BY-SA 4.0 or CC0	381	6,474,287
heldout-papers-nc	The same window, CC BY-NC-SA 4.0	16	257,150

Tokens are counted with the K2-Horizon tokenizer, without special tokens.

Validated PR eval (pr-validated)

Each item shows the repository, a pull request (its description with test results removed) and the diff of one revision, either the merged one or an earlier one. The model answers whether the fix passes validation: A) passes or B) fails. Labels: 139 passes, 15 fails. 15 upstream repositories; PRs merged 2026-10-16 to 2026-11-07.

Every label has evidence from outside the PR author:

• Tier B (117 items): CI passing on the merged head, approval by a human who is not the author, or reproduction or measurement by a non-author. A fail label comes from a CI failure, a breakage that got the change reverted, or a reviewer's rerun that still fails.
• Tier A-upstream (37 items): we reran the PR's own test. A final revision passes when the test fails twice at the merge base and passes twice at the merged head (33 items). An earlier revision of a PR fails when the test fails twice at that revision while the PR head passes (4 items).

Files:

• eval.jsonl: messages (system prompt, user prompt, reference answer).
• meta.jsonl: one line per item, same order as eval.jsonl: answer, label, tier, task (fix), repo, pr, url, merged_at, evidence (kind, source, author role, quote), component, prompt_tokens, diff_truncated, status.
• nll.jsonl: the full text of each PR for held-out NLL, with uid, tier, repo, pr, tokens, status.
• prs.jsonl: one line per PR unit with its items, label, evidence, merge and head SHAs, files and near-duplicate score.
• holdout_keys.json: what a later training build must drop to keep this eval clean: every PR or issue in an eval component, any document naming an eval PR number or merge or head SHA, any PR or commit by the same author in the same repository within the listed window that touches the same files, and any document whose changed lines reach MinHash Jaccard 0.5 with an eval PR. The MinHash signatures are included.
• MANIFEST.json and SHA256SUMS.

Version 3 retired 29 items, 23 NLL PRs and 29 PR units that came from a private source rather than upstream PR validation. They are not in this copy. Every line here has status active.

Versions 1 and 2 of the eval are not published. Version 3 kept every upstream item of version 2.

Held-out sets

heldout-prs, heldout-commits and heldout-papers are time splits: they begin the day after the training data ends (PRs and commits merged from 2026-10-16, papers dated from 2026-10-01). They use the same record formats as the training subsets of Avifenesh/k2h-corpus-domain-clean:

• heldout-prs: {"text"}, one merged pull request with description, labels, changed files, patch and review discussion (participants shown as author and reviewer N). 1,374 of 4,179 records include the patch; the others keep the description and discussion.
• heldout-commits: {"text"}, one commit with Repository:, Commit:, Date:, the message, file summary and patch (patch in 1,088 of 1,349).
• heldout-papers, heldout-papers-nc: {"text", "arxiv", "license", "license_url", "method"}, one paper converted from LaTeX.

The training PR and commit subsets were deduplicated against these sets with MinHash.

Licenses

• PRs and commits: each record names its repository. All repositories are open source: Apache-2.0 (vLLM, SGLang, FlashInfer, TensorRT-LLM, LMCache, Mooncake, speculators, Dynamo, TensorRT Model Optimizer, llm-compressor, TransformerEngine, LMDeploy, LightLLM, xLLM, xgrammar, ktransformers, transformers, DeepSpeed), MIT (llama.cpp, Triton, TileLang, flash-linear-attention, exllamav3, AITER, FlashMLA), BSD-3-Clause (PyTorch, ao, flash-attention, CUTLASS) and BSD-2-Clause (Liger-Kernel). Patches are under the repository's license. PR descriptions and review comments are public GitHub content by their authors; e-mail addresses are replaced with <email>.
• Papers: each record carries its arXiv license. heldout-papers is CC BY 4.0, CC BY-SA 4.0 or CC0 1.0. heldout-papers-nc is CC BY-NC-SA 4.0, non-commercial use only. Papers under the arXiv default license (274) or CC BY-NC-ND (25) are not included, because those licenses do not allow redistributing an adapted text. heldout-papers/LICENSE_AUDIT.json lists them by arXiv id.
• The eval prompts, labels and metadata we wrote are CC BY 4.0.

If you are an author of a pull request in these sets and want it removed, open a discussion on this dataset.

Known caveats

• pr-validated is small and skewed: 139 of 154 items are passes. Report accuracy per class, or balanced accuracy.
• Tier A-upstream reruns ran on one machine. A flaky test can flip a rerun label; each label needed two failing and two passing runs.
• Diffs longer than the prompt budget are truncated (diff_truncated in meta.jsonl).