Datasets:
The dataset viewer is not available for this subset.
Exception: FileNotFoundError
Message: Couldn't find any data file at /src/services/worker/cmunley/mixed-skills-env-v2b. Couldn't find 'cmunley/mixed-skills-env-v2b' on the Hugging Face Hub either: LocalEntryNotFoundError: An error happened while trying to locate the file on the Hub and we cannot find the requested files in the local cache. Please check your connection and try again or make sure your Internet connection is on.
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 268, in get_dataset_config_info
builder = load_dataset_builder(
path,
...<6 lines>...
**config_kwargs,
)
File "/src/services/worker/src/worker/utils.py", line 386, in safe_load_dataset_builder
dataset_module = dataset_module_factory(
repo_dir,
revision=revision,
download_config=download_config,
)
File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1178, in dataset_module_factory
raise FileNotFoundError(
...<2 lines>...
) from None
FileNotFoundError: Couldn't find any data file at /src/services/worker/cmunley/mixed-skills-env-v2b. Couldn't find 'cmunley/mixed-skills-env-v2b' on the Hugging Face Hub either: LocalEntryNotFoundError: An error happened while trying to locate the file on the Hub and we cannot find the requested files in the local cache. Please check your connection and try again or make sure your Internet connection is on.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Mixed Skills v2b: eight-domain verified workspace tasks
One OpenEnv server, 50 deterministic procedurally generated tasks across the eight evaluation domains: software engineering, industrial and physical systems, natural science, office work, finance and economics, math and formal reasoning, cybersecurity, and media production. Each task gives the agent a scratch workspace with input files and a written specification. The agent works with shell commands and is scored by a hidden verifier on the files it leaves behind.
tasks.jsonl lists every task id with its domain and the exact first-turn prompt. The image is
docker.io/chrismun/mixed-env:v2b (linux/amd64); it contains 102 task ids and the submission names the 50 used for training.
Interaction
Each turn the model replies with exactly one JSON object:
{"command": "<shell command>"}runs in the workspace (python3 with numpy, pandas and Pillow). At most 12 commands per episode; each command's output is truncated to its last 1500 characters, so strong solutions inspect withheadand do the work in a script.{"answer": "done"}ends the episode and scores the workspace.
Observations carry the command output, the commands left and, on most tasks, a one-line reply-format reminder. Replies that are not a single JSON object end the episode, which is then scored on the workspace as it stands.
Reward
Terminal and in [0, 1]. Every verifier is a weighted set of independent sub-checks recomputed from the task inputs, so partial progress earns partial reward and only a fully correct deliverable scores 1.0. References live only inside the image, agent commands run as an unprivileged user with memory and file-size limits, and verifiers that execute agent code do so in a subprocess with a timeout. An untouched workspace scores at most 0.25 on every task.
Calibration
Tasks were selected by measuring Qwen3.8-27B (thinking off, strict one-JSON-per-turn replies) with 8 to 16 rollouts per task, keeping tasks where the model sometimes fully solves the task but usually does not, so groups of rollouts carry reward variance. The sampling temperature was matched to the arena by replaying a known run: group variance on a saturated reference environment is reproduced at temperature 0.5 to 0.7, not 1.0, so selection used T=0.6.
- Final 50 tasks, fresh passes at T=0.6 (16 rollouts per task, 16 to 24 for the re-rolled ones): mean pass@1 0.66, estimated probability that a group of four rollouts has reward variance 0.81; selected from a larger pool on separate rollouts, with a cross-fit estimate of 0.78 for the selection procedure.
- Robustness to sampling temperature, fresh 8-rollout passes the selection never saw: 0.77 at T=0.5 and 0.80 at T=0.7 for the same probability.
- Prompt text: the task statements and the three global strings (preface, reply-format reminder, per-step hint) were rewritten by Qwen3.8-27B from the original
statements under strict gates (every number, file name, identifier, quoted string and JSON example must survive verbatim).
prompt_provenance.jsonrecords the model, attempts and the selection of rewrites. Difficulty was matched against the original wording by re-rolling the eight tasks whose pass@1 moved; one task,office-5, stays harder than in the original wording (0.20 against 0.49 at T=0.6). - A reply-format reminder in the prompt and a per-step hint raise strict mean reward (0.37 without either, 0.57 with both in an ablation), and they also raise reward variance per group. Nine tasks run bare, with no reminder and no hint, because they keep strong variance without it and so exercise unprompted strict-JSON replies.
- No transcript in the profiling runs touched verifier or task source.
Domains
| Domain | Skills exercised |
|---|---|
| software | multi-file bug fixing against hidden tests, implementing modules from a spec, CLI conversion, migration scripts, simulator debugging, performance and dependency problems |
| industrial | sensor log analysis and fault reports, control tuning by simulation, pipe and pump sizing, tariff billing, inventory policy, flow-line simulation |
| science | curve and rate fitting, sequence analysis and protein mass, ODE integration, hypothesis tests, titration equilibria, spectral analysis |
| office | payables reconciliation, meeting-minutes extraction, room scheduling, expense audits, payroll, contact deduplication |
| finance | amortization, market-data analytics, ledger audit, bond pricing, tax lots, FX arbitrage |
| math | exact number theory, proofs by computation, assignment optimization, shortest routes, constructive codes, absorbing Markov chains, impartial games |
| security | log incident analysis, hardening vulnerable code, layered-encoding and crypto forensics, detection rules, advisory matching (all synthetic, defensive) |
| media | subtitle retiming, image pipelines, audio synthesis, SVG charts, timecode conformance, markdown to RSS |
Files
tasks.jsonl:task_id,domain,promptfor each of the 50 training tasks, in training order (domains interleaved).
- Downloads last month
- 80