Ely2ba/reasoning-durability

Fine Until Fine-Tuned: model outputs

The texts the models wrote in the experiments of Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragile (Ely Sheikh, 2026, arXiv:2609.33559). With them you can read every response and training text behind the paper's numbers, re-score them, and re-run an arm on the same texts.

Dataset: Ely2ba/reasoning-durability. The code, the frozen inputs and the per-sample scores are in https://github.com/ely2ba/reasoning-durability. That repository's outputs/ tables have one row per sample and no text; the texts behind those rows are here.

Layout

The files are 1.97 GB compressed (6.96 GB as the runs wrote them). Paths mirror the original runs tree, runs/<family>/<run>/..., which is where the repository's src/collect.py and runners look. MANIFEST.csv lists every file with its rows, bytes and SHA-256, plus the files not included. TOTALS.csv gives totals per run. build_bundle.py made the bundle from the runs.

Path	Contents	Files	Rows
runs/{decision-probe,repetition-probe}/<run>/<label>/u<k>/<kind>.jsonl.gz, runs/anchor/anchor-20260924/<label>/u<k>/<kind>.jsonl.gz	TCES (arithmetic search) probe samples; kind is decision, think, competence or unforced	466	524,160
runs/endpoint-clone/endpoint-clone-20260910T162817Z/acquisition/corpus/group-NNNN/samples.json.gz	Clone corpus: the RL teacher's 8 samples on each of 800 TCES prompts; the clone S and the TCES arms train on it	800	6,400
runs/anchor/anchor-20260924/anchor.jsonl.gz, anchor.partial.jsonl.gz	M0's own TCES samples: the 4,480 the anchor arms train on, and all 4,800 drawn	2	9,280
runs/math/<run>/eval/<bench>/<label>/u<k>-<mode>.jsonl.gz	Evaluation samples on MATH-500 levels 3–5 and AIME 2025–2026	219	152,408
runs/math/pilot-20260924/reforce/<rule>/<label>/u<k>-forced.jsonl.gz	Forced samples continued with "Wait"	18	15,912
runs/math/<run>/rft.jsonl.gz	The training pool screen: one forced solution per screened problem. The correct ones, in arms.json order, are the arms' training texts	4	25,088
runs/math/<run>/p_samples.jsonl.gz	The P arm's training texts: up to 8 solutions per problem	3	1,737
runs/math/teacher-20260924/teacher/traces.jsonl.gz	gpt-oss-120b traces (8,480 on 4,480 problems) that the D-T, O-T and P-T students train on	1	4,480
runs/math/revision-sh-20260926/sharp-t05.jsonl.gz	O-sharp's training texts: the base's solutions at temperature 0.5	1	4,480
runs/math/revision-rl-20260926/habit.jsonl.gz	The habit control's training texts: the base's responses to chat prompts (see below)	1	519
runs/math/<run>/depth/<label>/texts.jsonl.gz	The forced MATH-500 texts, from the base and from each state, that are scored for the divergence	21	4,641
runs/**/*.npz	Divergence arrays: each fixed text scored per token under a base and a state	156	34,091 texts
runs/skill/skill-20260926/*/*.jsonl.gz	The skill part SK: base-7 multiplication and word sorting	19	13,360
runs/math/<run>/{arms,d2,replayed}.json, runs/skill/skill-20260926/{arms,calibration}.json	Training order, ids only	10	–

The models are Qwen3.5-35B-A3B-Base for main-35b-20260924, NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 for main-nemotron-20260924, gpt-oss-120b for the teacher traces, and Qwen3.5-9B-Base (or a LoRA fine-tune of it, named by label) for everything else. The code repository's records/ describes each run, and its outputs/README.md defines the labels and the score columns.

Fields

• Probe, evaluation, reforce and skill rows carry task_id and draw, the text as text (and, for the math and skill rows, token_ids), and the scores that outputs/README.md defines (correct, tokens, capped, …).
• The training files keep the model's token ids and add the decoded text beside them: rft, sharp-t05, habit and the anchor files have generation.completion_token_ids and generation.completion_text; p_samples has samples and sample_texts; depth texts have token_ids and text. Decoding follows the repository's common.models.Family: Qwen3.5-9B-Base's tokenizer (which the 35B shares) for the Qwen runs, and Nemotron 3 Nano's for main-nemotron-20260924. Completions in rft, p_samples, sharp-t05 and habit begin with the forced reasoning prefix (\n<think>\n for Qwen, <think>\n for Nemotron); in the anchor files the prefix ends the prompt instead.
• Clone corpus groups: records holds 8 samples. Each sample has prompt_token_ids (our TCES prompt), generation (completion_text, completion_token_ids, completion_logprobs, sample_id, sampling settings, reward) and reward (the verifier's result). The ids in the repository's data/tces/subsets.json are generation.sample_id.
• Teacher traces: task_id, plus kept, a list of traces, each with analysis (the reasoning), final (the answer) and teacher_tokens.
• .npz: task_ids, and per-position arrays tokens, lp, top and top_lp (NaN or -1 where unscored). src/collect.py (depth_pairs) gives the pairs of files that are compared.

Joining to the repository's outputs/

outputs/ table	File here	Key
math_samples.csv.gz	runs/math/<run>/eval/<bench>/<label>/<when>-<mode>.jsonl.gz, or reforce/<rule>/... when rule is not first	task_id, draw
tces_samples.csv.gz	runs/<decision-probe, repetition-probe or anchor>/<run>/<label>/u<update>/<kind>.jsonl.gz	task_id, draw
skill_samples.csv.gz	runs/skill/<run>/<set folder>/<setting or label>.jsonl.gz	task_id, draw
math_rft.csv.gz, math_p_samples.csv.gz	rft.jsonl.gz, p_samples.jsonl.gz	task_id
depth_texts.csv.gz	the .npz pair named by run, set, base, state	task_id

The row counts match those tables: 524,160 TCES rows, 168,320 math rows, 13,360 skill rows, 25,088 rft rows and 579 P rows. To re-run an arm, decompress a run folder into the repository's runs/ (the clone corpus goes under release/clone-corpus/), restore the math prompts as described below, and follow the repository's README.

Removed, withheld and not included

• Third-party prompts. prompt_token_ids is removed from rft, p_samples, sharp-t05 and habit. It is the tokenized NuminaMath-TIR problem or Tulu 3 chat prompt. The repository rebuilds the prompts from their public sources by id (make inputs); Family.render then reproduces the removed tokens exactly (math: f.render(maths.INSTRUCTION + problem); habit: f.render(prompt)). Our own TCES prompts are kept.
• Habit responses to jailbreak and safety prompts. The habit control used the base's responses to 650 chat prompts. The 131 responses to WildJailbreak (73) and WildGuardMix (58) prompts are withheld and available on request. The other 519 are included. habit_withheld.jsonl.gz lists the withheld rows by position, id and source, without text, so the file can be completed.
• Private identifiers. Tinker checkpoint URIs, which carry account session ids, are removed from the clone corpus. Some model texts contain strings that look like identifiers, such as UUIDs in a degenerate sample or example e-mail addresses in code. They are model output.
• Not included, available on request: the LoRA adapters of the endpoint-clone run (19, 7.19 GB, listed in MANIFEST.csv with SHA-256); the 86 Tinker training states (data/checkpoints.json in the repository); the RL teacher T's training rollouts and the clone S's in-training probes (endpoint-clone run); the runs of the original replay study (replay-v1), which trained the RP and RS states that the TCES probes use; the aborted probe run dc-20260923; and the per-update training logs (summarized in outputs/train_log.csv.gz).

Terms

Each model's outputs are under that model's terms. Our synthetic tasks, the collection and its metadata are under Apache-2.0. See LICENSE.md.

Citation

@misc{sheikh2026fine,
  title         = {Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragile},
  author        = {Sheikh, Ely},
  year          = {2026},
  eprint        = {2609.33559},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2609.33559}
}