splits/README.md

Splits

Training split only. This repository publishes the dev_task_names half of each pinned split. The held-out/test tasks are not included, and neither are the published splits that are not the pin. Every split file here is reproduced byte-for-byte from the source repository, so assignment_sha256 still verifies and heldout_task_names still names the absent test tasks.

One folder per dataset. The file inside is the pinned resolved train/test assignment over the tasks of that dataset, published without its private random seed.

The tasks themselves stay in the dataset folder at the top level of this repository, so the dev_task_names listed in splits/tau3-bench/tau3-banking-70-s001.json are directories under tau3-bench/tasks/. The heldout_task_names are not.

splits/
  QuantitativeFinance-Bench/
  healthbench/
  hle/
  tau3-bench/

Which split we use

Several datasets carry more than one published split. Exactly one of them is the pin: the split that the skilltrainbench launch skills hardcode and that the reported results were produced on. Only the pin is published here; the others stay in the source repository, for replication and for pipeline tests, and a number produced on them is not comparable to a number produced on the pin.

Dataset	Pinned split	Task set	Dev / train	Held-out / test	Learner pool	Claude Code max_turns	OpenHands max_iterations
QuantitativeFinance-Bench	qf-medhard-strat-80-s005	medium_hard	54	14	160M	41	45
healthbench	hb-hardbeh-s002	hard_behavioural	200	160	80M	11	12
hle	hle-stem517-s001	stem517	412	105	400M	36	40
tau3-bench	tau3-banking-70-s001	banking_knowledge	67	30	360M	51	60

The four benchmark pools sum to 1,000M and that total is a hard limit. Each pool is dev x 4 x the mean weighted tokens (input + 5 x output) of a within-cutoff GLM 5.3 Flash attempt, the four scaled together to sum to 1B, which is what a two-person hackathon team's $200 of GLM buys. The mean counts an over-cutoff attempt with the tokens it had reached at the cut rather than dropping it.

The learner pool is the --budget-tokens pool for skill formation. It caps the learner only. Curator spend, judge spend and simulator spend are metered in their own pools and are never charged against it.

The tool-call cutoff is one number per benchmark, applied identically to every learner. Claude Code max_turns counts model responses and OpenHands max_iterations counts loop steps, so neither is an exact tool-call ceiling; the numbers above are the serial-call allowance, with OpenHands given about 20% more. A cutoff hit is a failed attempt, and its rate is reported per arm next to the lift.

The published results predate this table. They were measured at 100M (QuantitativeFinance-Bench), 100M (healthbench), 300M (tau3-bench) and 500M (hle), with healthbench at 4 turns and the others effectively uncapped. Those runs stay valid as published, but a new run and an old run are not on the same budget. Rescoring an old run under the new cutoffs is not valid either: the probe signal, and so the skill, would have differed.

Each pinned split verifies against these digests:

Dataset	assignment_sha256
QuantitativeFinance-Bench	f256dfa217a335f88a925c088b78289553de0a9cadaec96153d682cabf60ec0f
healthbench	32b8ab7615506a55e9fa3ca89ee1d6e68f66a771f62636f2876d2ec99ca68cfb
tau3-bench	90d6bcc7f58b181725dc4c0f21960b1459c6b9bacdb6ef09f2fb7eda30fb27b6
hle	3b1755f035bf7135bfdd4b40bf376cb157cc7f7a49739e48da92e35be06d5aeb

Pins that are not splits

A run is only reproducible if the environment is pinned too. These values live in src/skilltrainbench/harbor_policy.py in the skilltrainbench repository and are applied when a task tree is staged, so a task Dockerfile in this repository may name an unpinned tag while the run that used it did not.

Dataset	Base image	Upstream code pinned at build time
QuantitativeFinance-Bench	doenermannti/finance-bench-sandbox@sha256:a6766c8b32cae3425feaddb8d9bac185de866323307290b7c74aed30f9c928ef	none
healthbench	python:3.12-slim@sha256:2fe5997d249a808b8eeea52c58a1dbffbba28754dc11699ef5c029f2d818ce79	openai/simple-evals@496e4570446e9f0499a00557a231a440c446df76
tau3-bench	python:3.12-slim@sha256:2fe5997d249a808b8eeea52c58a1dbffbba28754dc11699ef5c029f2d818ce79	sierra-research/tau2-bench@672227c6b6
hle	python:3.11-slim@sha256:9534e5a8e315485d4061ed659af0fd78a284c015f9b73661b41d6bab25604534	harbor-framework/harbor-datasets@37db108843a49bb31a592e37a75e2c40dc3f9749

The tau3-bench images clone tau2-bench while they build, so that clone has to be pinned or the image silently tracks upstream. Override it with STBENCH_TAU3BENCH_TAU2_COMMIT only when you mean to.

Grader and simulator models

These are configuration, not dataset content, and a run records what it used.

• healthbench. The grader is the pinned simple-evals HealthBench grader, vendored as healthbench_eval.py with SHA-256 b763d8c1f53ecd165f0d170a32bc948216b3dfeadafe238b2083328e1af4b2df. The official judge configuration for the benchmark is gpt-4.1-2025-04-14 at temperature 0.5. The reported hb-hardbeh-s002 run deliberately deviates and used claude-haiku-4-5-20251001 at temperature 0, with its own 40M judge pool. Say which one you used.
• tau3-bench. The simulated user and the natural-language assertion judge run through TAU2_USER_MODEL and TAU2_NL_ASSERTIONS_MODEL, both openai/claude-sonnet-4-6 with TAU2_USER_REASONING_EFFORT=low in the reported run. They are metered in a separate simulator pool.

Selecting a split

--benchmark tau3bench --task-set banking_knowledge --split-id tau3-banking-70-s001

dev_task_names is the training set and heldout_task_names is the test set. Split identifiers are immutable. If a task set definition or the task bytes change, create a new identifier rather than reusing one; the loader fails closed when task_set_definition_sha256 or selected_source_snapshot_sha256 no longer matches.

Fields

• schema_version: split artifact schema.
• split_id: immutable identifier for this assignment.
• benchmark: skilltrainbench benchmark adapter.
• task_set: selected task-set identifier.
• train_ratio_in_split: requested train proportion.
• stratified: whether the assignment was stratified.
• dev_task_names: development names; treat these as train tasks.
• heldout_task_names: held-out names; treat these as test tasks.
• counts: resolved development and held-out counts.
• strata: counts per stratum, where the assignment was stratified.
• task_set_definition_sha256: digest of the task-set definition.
• selected_source_snapshot_sha256: digest of the selected task bytes used to resolve this assignment.
• assignment_sha256: digest of the public, seed-stripped assignment.

The private random seed is removed from every file published here. It stays in the skilltrainbench repository under the ignored private/splits/ directory.

Download

hf download armin-aptura/skilltrainbench-public --repo-type dataset \
  --include "splits/**" --local-dir .

For one dataset and its split together:

hf download armin-aptura/skilltrainbench-public --repo-type dataset \
  --include "tau3-bench/**" --include "splits/tau3-bench/**" --local-dir .