README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: tmax_clean
        path: data/tmax_clean.parquet
      - split: tmax_repair
        path: data/tmax_repair.parquet
      - split: tmax_killed
        path: data/tmax_killed.parquet
      - split: calibforge_clean
        path: data/calibforge_clean.parquet
      - split: calibforge_repair
        path: data/calibforge_repair.parquet
      - split: calibforge_killed
        path: data/calibforge_killed.parquet
      - split: tmax_river_clean
        path: data/tmax_river_clean.parquet
      - split: tmax_river_noanswer
        path: data/tmax_river_noanswer.parquet
task_categories:
  - reinforcement-learning
language:
  - en
tags:
  - terminal-bench
  - agentic-rl
  - verifier-audit
size_categories:
  - 10K<n<100K

tb-full-data

Every TMAX and CalibForge task we have reviewed with one of our prompt rubrics, in one dataset. Each task appears exactly once, with its most recent verdict, where it came from (origin), every earlier verdict (verdict_history), and, where one exists, a solution verified the way the line that produced it verified it. 20,025 tasks in 6 splits: clean, repair and killed for each corpus.

split	rows	with a solution	security-domain rows
tmax_clean	6,085	6,085	542
tmax_repair	7,454	2,778	874
tmax_killed	1,055	3	106
calibforge_clean	3,294	3,294	285
calibforge_repair	1,998	1,266	175
calibforge_killed	139	82	9
tmax_river_clean	3,186	3,186	276
tmax_river_noanswer	130	0	24

from datasets import load_dataset
ds = load_dataset("Fzz1/tb-full-data")   # ds["tmax_clean"], ds["tmax_repair"], ds["calibforge_clean"], ...

Correction (2026-10-24): five rows moved to repair

A post-publish verifier audit of Fzz1/tb-training-mix-agentpick found five of its rows, also here in tmax_clean (1) and calibforge_clean (4), whose verifier can refuse a correct solution, can be overridden by the agent, or can fail for a reason unrelated to the solution. They move to tmax_repair / calibforge_repair with tier FLAGGED, tier_source / verdict_source post_publish_verifier_audit_20261024, and the stage appended to verdict_history and origin; their solutions and every other column are kept. None of them is in tmax_river_clean.

task	from -> to	verdict	defect
task_000283_b8c17c0d	tmax_clean -> tmax_repair	INSTR-VERIFIER-MISMATCH	protected file the instruction tells you to create
model-training_20260504_125558_004	calibforge_clean -> calibforge_repair	INSTR-VERIFIER-MISMATCH	grader prefers agent-writable /workspace/tests
software-engineering_20260419_112101_010	calibforge_clean -> calibforge_repair	INSTR-VERIFIER-MISMATCH	grader prefers agent-writable /workspace/tests
software-engineering_20260510_130518_011	calibforge_clean -> calibforge_repair	INSTR-VERIFIER-MISMATCH	grader prefers agent-writable /workspace/tests
data-querying_20260617_104707_001	calibforge_clean -> calibforge_repair	UNSTABLE-REWARD	silent grade-time rdflib install

Counts: tmax_clean 6,086 -> 6,085, tmax_repair 7,453 -> 7,454, calibforge_clean 3,298 -> 3,294, calibforge_repair 1,994 -> 1,998; every other split, row and package is unchanged.

The rule

• <corpus>_clean: tier SEED, verdict CLEAN, and a verified solution (solution_replay_reward 1.0).
• <corpus>_repair: every other judged task that no earlier review killed. None is held back.
• <corpus>_killed: kept for audit, NOT training data. Every task an earlier review or gate of ours removed: TMAX 1,055 (first removed by: v1r2 555, v13 tb-training-mix-agentpick 203, v13 tb-training-mix 129, gate2 129, gate1 23, v12 8, v8 8), CalibForge 139 (answer leaks in Fzz1/tb-training-mix-agentpick's review). The stage, the label it recorded and the reason are in origin, verdict_history and the split file.
• Answer leaks found by the latest review are in repair, with the verdict visible (TMAX 3,332, CalibForge 557). Filter on verdict before training on a repair split: ds["tmax_repair"].filter(lambda r: r["verdict"] != "ANSWER-LEAK").
• One verdict per task, the most recent of ours. Earlier verdicts and kills are in verdict_history and in the split files (metadata/split_assignment_<corpus>.tsv).
• Nothing was re-judged or re-rolled for this release. Rows come from the lines that produced them, as those lines published or recorded them; origin names the line, dataset and revision of every row.
• A verified solution means verified by its own line:
  • this dataset's own TMAX and CalibForge lines: one replay of the first reward-1 attempt in a fresh sandbox graded 1.0 (solution_source t14_replay_verified, and kimi_max_replay_verified for the TMAX Kimi-max retries);
  • rows from Fzz1/tb-training-mix-agentpick: its gate-2 replay of an agent attempt (TMAX, agentpick_replay_verified) or a strict majority of graded replays (CalibForge, agentpick_cf_replay_majority);
  • 24 TMAX rows from Fzz1/Tmax-Tasks-Clean reaudit: its published solve.sh, carried only where that release's instruction, setup and tests are byte-identical to upstream, it adds no verifier file and runs no pre-test hook (ttc_reaudit_gold:identical_verifier).
• metadata/killed_tasks.tsv lists every kill: task id, upstream and revision, the killing stage, the label as recorded, the reason, and the record it comes from in our working repository (record_file, the path with internal directory prefixes removed, and record_commit). It also lists, as superseded, 726 TMAX tasks an early stage removed and a later review or release took back; those are in the split their later verdict gives.

TMAX

allenai/tmax-15k-open-instruct at 7b090eca: 14,594 of its 14,601 tasks.

• The tasks of this dataset's TMAX line: judged by Kimi K3 under rubric v13 (some earlier v13 verdicts), rolled out by DeepSeek V4.1 Flash (effort high, retries at max), replayed once.
• 977 SEED/CLEAN tasks of this line that DeepSeek left unsolved after 3 attempts were rolled again on Kimi K3 (effort max, top_p 0.95, up to 3 attempts, stopping at the first reward 1, then one deciding replay). 140 were solved and replay-verified and are in tmax_clean (solution_model moonshotai/Kimi-K3, solution_source kimi_max_replay_verified, rollout_best_attempt kimi_max_attempt<k>; rollout_attempts counts both models' attempts, and the DeepSeek reward columns keep the DeepSeek attempts). 837 remain in repair, unsolved after 3 DeepSeek + 3 Kimi attempts.
• The tasks earlier lines had already reviewed or released: in Fzz1/Tmax-Tasks-Clean, Fzz1/tb-training-mix-agentpick or Fzz1/tb-training-mix, killed by an earlier review or gate, or reviewed under an earlier rubric. Where a v13 verdict exists it decides. Otherwise the task's latest pre-v13 verdict is kept in its own vocabulary (v8 1,107, v1r1_review 163, v2 160, audit452 97, v3 70, v1r1_rejudge 49, v1r2_fable 4, v1r2_opus 4, v4_disp 3, v7 1, v6_b1 1): rubric and tier_source name the stage, verdict holds its label, tier is empty, and the row is in repair (pre_v13_verdict_only), or in tmax_killed if an earlier stage killed it. These earlier rubrics are not mapped to v13 tiers because they disagree with v13 too often to be read as one: of tasks judged by both, v8 PASS became v13 SEED on 40 of 57.
• A solution an earlier line verified against different verifier bytes is not carried (most of Fzz1/Tmax-Tasks-Clean's reaudit and longlongcheck solutions); those tasks are in repair unless another verified solution exists.
• Not here: 7 tasks with no prompt review on record (listed by a release with no verdict anywhere, including tasks the 452-corpus audit held without reviewing them); each is in the split file with the reason.
• audit452 is the 452-corpus static semantic audit of Fzz1/Tmax-Tasks-Clean's reaudit corpus (protocol 2.2, independent A/B reviews with an adjudicating C on disagreement); its label is its assessment (STATIC_RISK, SUPPORTED_WITHIN_SCOPE). It killed no task.

CalibForge

AweAI-Team/CalibForge at fb1e7544: all 5,431 tasks.

• 4,665 from this dataset's CalibForge line: judged by Kimi K3 (v13-harbor3); pass 1 on DeepSeek V4.1 Flash (high, top_p 1.0) for 544 tasks and Kimi K3 (low, top_p 0.95) for the rest; retries on Kimi K3 (max, top_p 0.95).
• 766 from Fzz1/tb-training-mix-agentpick: judged by DeepSeek V4.1-Flash (max), GLM-5.3 (low) where no valid V4.1 row exists; DeepSeek Flash rollouts. Its 139 answer-leak kills are in calibforge_killed.
• 38 tasks that reuse a Terminal-Bench 2.1 task's test scaffold stay in calibforge_repair (reason "TB2.1 sibling (copied test scaffold)"), as in that release.

Not in this dataset

Other corpora we reviewed or curated are in their own datasets and not repeated here: RST, TerminalWorld, Terminal-Lego and Xiaomi MiMo (Fzz1/tb-training-mix-agentpick, Fzz1/tb-training-mix), and Fzz1/SWE-Smith-Seeds-Clean, Fzz1/SWE-Rebench-Tasks-Clean, Fzz1/TB2-Vendor-38.

tmax_river_clean and tmax_river_noanswer (the "river" list)

In the default config, with the same 122 columns as every split. tmax_river_clean: every task of the "river" list, a user-supplied list of 3,316 TMAX task ids (metadata/tmax_river_list.tsv, where in_split = 1 marks the ids in tmax_river_clean; input list sha256 0d9da26cd6cb94922fc5e220a6901df76075b48fbcc45d6128794341255d4766) that carries a replay-verified solution: 3,186 rows, 276 security-domain. The rows are copies: each task also stays in its own split, named in source_split (tmax_clean 2,125, tmax_repair 950, tmax_killed 110, excluded 1). The 111 tmax_killed / excluded rows are included by request; their verdicts are unchanged and they are not training data under this card's rule.

Solutions, each verified by one fresh-sandbox replay of the solving attempt graded 1.0, with sha256 = the replayed actions: this dataset's TMAX line (DeepSeek V4.1 Flash, the Kimi-max retries), and two later lines on this list: nr409 (Kimi K3, effort low then max on retries, top_p 0.95; rows never rolled before) and k254 (Kimi K3, effort max on all 3 attempts, top_p 0.95; the rows still without a solution). solution_source: t14_replay_verified 2,530, nr409_replay_verified 404, k254_replay_verified 161, agentpick_replay_verified 51, ttc_reaudit_gold 26, kimi_max_replay_verified 13, t14_rereplay_verified 1. Tier and verdict are unchanged; 93 tmax_repair rows have a v13 SEED/CLEAN verdict and a solution found by the later lines, i.e. would qualify for clean (newly_solved_v13_seed_clean). Per-row extras live in metadata/tmax_river_list.tsv, one line per listed id: source_split, list_index (the id's line in the input list), reason (from metadata/split_assignment_tmax.tsv), nr409_final, k254_final, newly_solved_v13_seed_clean. For rows a later line rolled, rollout_* / replay_* describe those attempts (rows rolled before keep their earlier attempts and add the new ones). The ids without a verified solution are listed in the metadata file with the reason.

tmax_river_noanswer holds the other 130 ids of the list: no attempt produced a replay-verified solution, so solution is empty, and rollout_* / replay_* keep the attempts as recorded. 92 were unsolved after Kimi K3 max x3 (k254; earlier attempts, if any, are in the row), 37 after DeepSeek x3 and Kimi K3 max x3 (the Kimi-max line), and 1 solved once by k254 but its deciding replay graded 0. 24 are security-domain; by source_split: tmax_repair 114, tmax_killed 16. In metadata/tmax_river_list.tsv these lines read in split tmax_river_noanswer: <reason>.

Columns

The first 70 columns are Fzz1/tb-training-mix-agentpick's, with the same names, order and types; the next 50 are this dataset's rollout and replay columns; the last two are origin and verdict_history.

• origin: where the row and its verdict come from: this dataset's own line, or the earlier dataset and revision it was published in (or the split file that listed it without shipping it), the verdict's source, the solution's source, and, for a killed row, the stage that killed it.
• verdict_history: a JSON list of every earlier verdict and kill of the task, oldest first (stage, label, and where a line recorded one, its class). The split files carry the same.
• solution_source: how the shipped solution was verified (see the rule above); empty when there is no solution.
• The 50 rollout and replay columns describe this dataset's own TMAX and CalibForge rollouts (rollout_*, replay_*). They are null on every row this dataset did not roll out (all rows from earlier lines); rollout_attempts is 0 there. Those lines' own solution and sizing facts are in the first 70 columns as they published them.
• Rows from earlier releases keep those releases' values in the first 70 columns, including column_gaps; rows a release listed without shipping are filled from that release's own side tables and the package, and column_gaps names every column with no source.
• CalibForge rows mix models: solution_model (deepseek-flash, moonshotai/Kimi-K3) and solution_reasoning_effort are the replayed attempt's. The split file gives model, effort and top_p of every attempt of this dataset's CalibForge rollouts.

Packages

data/tasks-<split>.tar holds each task as tasks/<task_id>/, byte-identical to its upstream at the revisions above, except:

• TMAX: instruction.md, setup.sh, tests/test.sh, plus environment/Dockerfile (FROM <the task's image> with the setup script inlined as comments; equal to the dockerfile column);
• CalibForge rows with dockerfile_repaired true: environment/Dockerfile is a single FROM aweaiteam/calibforge@sha256:<digest> line (what the rollout ran).

Canary markers

The canary scan found 167 benchmark canary marker(s) across the uploaded files. They are kept as they are.