permissive-post-train/evaluation-records

MixtureVitae post-training evaluation records

The evaluation records behind the tables and figures of Strong Post-Training from Permissive, Reasoning-Dominant Pretraining (NeurIPS 2026), in one place. The pipeline that turns them into the paper's tables, figures and quoted numbers is mv-post-train. Its mv_post_train/inputs.json pins every file here by SHA-256 and by a revision of this repository, and mvpt fetch-inputs downloads them:

pip install -e . -r requirements-lock.txt
mvpt fetch-inputs        # every input of the paper, from here and its other sources
mvpt build --check       # rebuild every generated file and compare with results/

Some files are copies of public ones, and inputs.json keeps their original sources as fallbacks: R13 and R13-language of Ali-Elganzory/permissive-llms-artifacts at commit 8cc2e51, R14-scores and R14-ifeval of laion/mv-posttrain-eval-responses at revision e820fb34 (R14), and S2-0.4b and S2-1.7b of the two checkpoints' model repositories. The only other input of the paper is a small file in the pipeline repository itself, its setup values. docs/inputs.md of the pipeline describes every input.

Files

The IDs are the pipeline's input IDs. Table and figure numbers are those of the paper's revision of 2026-10-28, and for O1 to O7 those of 2026-11-01.

ID	File	Size	SHA-256	What it holds
R13	protocol_a/canonical_reasoning_per_sample.csv	105 MB	867ff6d6782b…	The main Protocol A records: one score per checkpoint, seed, task and question, for 57 checkpoints. Tables 1, 14, 16 to 18, 20, 21, 38 (its decontaminated rows) and 42 (its rows without masking), Figures 2 to 4.
R15	protocol_a/comma_reasoning_unified_per_sample.csv	12 MB	a8e33f163ce4…	Protocol A records of the front-loaded Comma0.1 arm (six checkpoints). Tables 17, 18, 20 and 21, Figures 2 to 4.
R18	protocol_a/ablation_decontamination_0_4B_reasoning_per_sample.csv	61 MB	021b81d039ec…	Protocol A records of the 36 checkpoints of the 100B mixture ablation, the MV model without decontamination and the 0.4B comparison. Tables 35, 37, 38 (its rows without decontamination) and 40.
R19	protocol_a/masked_reasoning_per_sample.csv	9 MB	8084115c57e8…	Protocol A records of the four masked MV checkpoints (copies of the laion/*-masked models with the same weights). Table 42's rows with masking.
R20	protocol_a/fineweb_edu_16k_april_base_unified_per_sample.csv	2 MB	153341e3f95a…	Protocol A records of the FineWeb-Edu 16K base that its SFT and DPO models were trained from. Tables 1, 14, 17, 18, 20 and 21, Figures 2 to 4.
R21	protocol_a/ifeval_strict_per_sample_protocol_a.csv	34 MB	eac0a09812bc…	Protocol A IFEval verdicts with the strict checker, one row per checkpoint, seed and instruction, for 112 checkpoints. Every Protocol A IFEval score of the paper except DCLM-16K's, which R24 holds.
R22	protocol_a/phase1_livecodebenchv5_per_sample.csv	17 MB	80b2d27da77a…	Protocol A LiveCodeBench v5 verdicts, one row per checkpoint, seed and problem, for 112 checkpoints, on the 369 problems of LiveCodeBench v5 that Phase 2 runs. Every Protocol A LiveCodeBench score of the paper except DCLM-16K's, which R24 holds.
R24	protocol_a/dclm_16k_reasoning_evals_per_sample.csv	6 MB	fcb6bd7b79f3…	Protocol A records of the DCLM 16K checkpoints (base, Tulu3 SFT and DPO), with their IFEval verdicts and their LiveCodeBench v5 verdicts (task LiveCodeBenchv5) next to v2, because R21 and R22 do not cover them. DCLM's 16K rows of Tables 1 and 14.
R13-language	language/ablation_frontloading_x_permissiveness_language_evaluations.csv	16 kB	e108272f2227…	One language score per task for the front-loading study's checkpoints without Comma0.1, and each task's shot count and metric. The shot counts and metrics of Table 13. The scores are a second source for most language cells of Tables 17 and 22, which print the runs of R8, R8-fineweb and R23.
R8	language/language_eval-logs.zip	2.9 GB	0492e8a69dd6…	LM Evaluation Harness results with per-question files and three prompt-example seeds, for 19 checkpoints of MV, Comma0.1 and Nemotron-CC-HQ. Tables 15, 17, 19, 22, 30 to 32, 39 and 43, and Figure 2's language panel.
R8-fineweb	language/logs-fineweb.zip	2.0 GB	d940f546b4dc…	The same for 7 FineWeb-Edu checkpoints. Tables 15, 17, 19, 22, 30 and 31, and Figure 2's language panel.
R16	language/language_eval_all_models_2026-10-16.zip	3 MB	bb0b7b1a2fc7…	LM Evaluation Harness results for 82 checkpoints, one prompt-example seed, without per-question files. Tables 15, 30, 36, 39, 41 and 43.
R17	language/masked_models_lang_evals.zip	0.5 MB	c53217be762c…	LM Evaluation Harness results for five masked checkpoints. Table 43's rows with masking.
R23	language/language_eval_all_models_3seed_2026-10-26.zip	3.5 GB	1a0422576760…	LM Evaluation Harness results with per-question files and three prompt-example seeds (1234, 42 and 2025) for 24 checkpoints, run on 2026-10-26, and one-seed results files for 41 more, which are R16's. The front-loading study's language cells that R8 does not hold, and the intervals of its language contrasts: Tables 17, 19 and 22, and Figure 2's language panel.
R14-scores	phase2/eval_scores.parquet	70 kB	666002c3c4e9…	The three-seed Phase 2 scores of R14, with IFEval under all four definitions. Every Phase 2 result of the paper.
R14-ifeval	phase2/ifeval_allmetrics_persample.parquet	0.2 MB	1dd7734ef79a…	R14's IFEval counts per prompt and seed: instructions, and those followed with the strict and the loose checker.
R14-meta	phase2/eval_responses_meta.parquet	2 MB	59872f7e7a31…	The small columns (model, stage, task, question, seed, correctness, checkpoint folder, token limit) of R14's answer file, eval_responses.parquet (10 GB), which is not copied here. mvpt fetch-phase2-records writes the same file from R14.
R14-lengths	phase2/answer_lengths.parquet	6 MB	9863758e52d1…	One row per Phase 2 answer of R14, without its text: its length, and whether the grader found an answer and a code block. mvpt measure-phase2-answers writes the same file from R14 in about two hours.
S1	snapshots/hf_checkpoints.json	173 kB	4d4723136230…	A snapshot of the Hugging Face records of every checkpoint the paper evaluates (repos, weight hashes, model cards, training summaries, context settings), taken on 2026-10-12 and 2026-10-13 and updated on 2026-10-31 with the Phase 2 checkpoints uploaded on 2026-10-30 and HuggingFaceTB's SmolLM2-1.7B-Instruct-16k. It gives each checkpoint's parent in the results manifest.
S2-0.4b, S2-1.7b	setup/*.config.json	1 kB each	24a64d86ef89…, b2d8faeee3c6…	The config.json of the 0.4B and 1.7B MV base models, the architectures of Table 5.
R6	older/logs_without_gpqa_items.zip	416 MB	2f9db28c0e7a…	LM Evaluation Harness and Evalchemy results files of the older evaluations, January to May 2026, without the GPQA Diamond items (see below). No table of the paper prints them, but the results manifest keeps them.
O1	overlap/excluded_items.csv	57 kB	742229278d1a…	The 1,720 benchmark items of 17 benchmarks that the audit of the training data identified, one row per benchmark and item ID. The comparisons without the identified items leave them out: Table 4 and Appendix A.3.
O2	overlap/overlap_sensitivity.csv	75 kB	623541e0a3f2…	Every comparison of Phase 2, Table 4, the front-loading study and the language suite, with all items and without O1's, each with its estimate, 95% interval and Benjamini-Hochberg q-value: 647 rows. Appendix A.3.
O3	overlap/found.parquet	1.5 MB	e2536a6f4036…	Where each item was found: the training dataset, its row and turn, the source within it, and the evidence (13-gram or exact). Appendix A.3's table of items found.
O4	overlap/mv_question_overlap_rates.csv	1 kB	2b7b2b45460b…	The share of the questions of the AceReason, Nemotron-Cascade and Superior-Reasoning 30K samples trained on that match MixtureVitae's pretraining reasoning data other than OpenThoughts3, by domain. Appendix E.
O5	overlap/mv_question_overlap_rates_ot3.csv	0.3 kB	966ea7ba4e6b…	The same for the Phase 2 OpenThoughts3 30K sample. Appendix E and §3.4.
O6	overlap/phase2_tokens.csv	1 kB	e9d8658a7418…	The training tokens of the four Phase 2 30K samples as trained: length per sample, the share cut at 16,384 tokens, and the supervised tokens. Appendix A.2.
O7	overlap/phase2_ot3_release.csv	0.4 kB	94e064d554d3…	The OpenThoughts3 30K sample as released, by domain: how many answers close their reasoning. Appendix A.2.

The full SHA-256 of each file is in inputs.json, which checks every download.

Formats

Protocol A records (R13, R15, R18 to R22, R24) have one row per checkpoint, decoding seed, benchmark and question: model_id, stage, model_group, front_loading, context_length, seed, benchmark, sample_id, correct, and in some files the experiments the checkpoint belongs to. They hold no question or answer text. Protocol A is greedy decoding with seeds 42, 1234 and 2025 (the paper's Table 11).

Language results (R8, R8-fineweb, R16, R17, R23) are the results files that the LM Evaluation Harness writes, zipped as received. R8, R8-fineweb and R23 also hold its per-question files, which contain the benchmark questions and the models' answers. R23 keeps each checkpoint, task and seed in a folder of its own (<checkpoint>/<task>/seed<N>/<checkpoint>/).

Older evaluations (R6) are the results files as received, except for GPQA Diamond. GPQA asks that its items not be published in plain text, and Evalchemy saves every item it scores with the results, so here each GPQA item keeps only its GPQA record ID, the letter of the correct choice, and the model's answer in each repetition if it is a letter (otherwise an empty string). The scores and every other file are unchanged. release/records/strip_gpqa.py of the pipeline writes this copy from the original.

The paths with /e/scratch/jureap59/ in the results files are the working directory of the cluster the evaluations ran on.

The benchmark-overlap audit (O1 to O7) looked for the paper's benchmark items in the training data with 13-gram and exact matching. Its files hold item IDs, row numbers, dataset and shard names and counts, and no benchmark or training text. The pipeline's docs/inputs.md describes each file, and mvpt render-overlap turns them into the paper's Table 4, Appendix A.3 and the overlap numbers of the text.

License

The records are under CC BY 4.0. The benchmark questions in R6, R8, R8-fineweb and R23 keep the licenses of their benchmarks, and the model answers in them those of the models that wrote them.

Versions

Records are added here as new runs and checkpoints are published. Each version of the pipeline pins the revision it read, so a release of the paper's tables always names the exact files behind it.