Datasets:
FineEnvs v1: 5,000 kept tasks
FineEnvs / SmolDataEnvs
Dataset stats
- Original tasks: 5,000
- Kept: 5,000
- Changed: 5,000
- Archived: 0
- Images: 1 image before, 1 image now.
Dataset characteristics
SmolDataEnvs (source revision 4719635555de) asks agents to inspect supplied datasets, compute an answer with Python data-analysis tools, and write it to /workdir/answer.txt.
Domains
Data Science: statistics, correlations, category frequencies, proportions and model evaluation. Data Processing: column conversion, missing values, filtering and aggregation of heterogeneous tabular data.
Capabilities
Exploration: inspect input files, shapes and column types before computing an answer. Specification Adherence: produce a single value, short label, yes/no answer or comma-separated list. The grader checks the final answer through exact, numeric, list and mathematical equivalence matching; it does not evaluate the analysis trace. Labels follow CLI-Universe.
Intended benchmark transfer: DS-1000 mini-200, through dataset inspection and use of pandas, NumPy and related analysis libraries. This is an expected transfer, not a measured training gain.
Per-stage breakdown
Counts can overlap because a task can require changes at several stages.
- 1. Build and setup: 5,000 changed, 0 archived. Changed 5,000 tasks to download inputs ahead of time and package them under
setup_files/inputs, avoiding Hugging Face rate limits during parallel setup. Upstream used the mutable:baseimage tag; we pinned the same image to an exact version. - 2. Reference solution (reward 1): 5,000 changed, 0 archived. Added
solution/solve.shto 5,000 tasks to run our reference solution checker; all passed. - 3. No-op baseline (reward 0): 5,000 changed, 0 archived. Changed 5,000 tasks to normalize the JSON score key from
correctnesstoreward; all recorded reward 0. The checker flagged missing grader execution for all tasks. Inspected a random sample of 50: all were valid no-ops, with the shell verifier writing zero and exiting before the Python grader started. Their reference trials confirmed that the grader could run successfully. - 4. Rule-based checks: 4,576 changed, 0 archived. Flagged 4,745 tasks for relative paths. The automatic resolver cleared warnings for 4,575 tasks; 170 still have warnings. Inspected a random sample of 50, and all warnings were harmless.
Git-history and timestamp checks were not applicable because no Git repositories were found at the scanned roots.
Per-task resource requirements
| Kept tasks | CPUs per task | Memory per task | Agent timeout |
|---|---|---|---|
| 4,932 | 1 | 1 GiB | 600 s |
| 3 | 1 | 1 GiB | 900 s |
| 6 | 2 | 4 GiB | 600 s |
| 59 | 2 | 4 GiB | 900 s |
All tasks have a 600 s setup budget and a 120 s verifier timeout. Source resource limits were preserved.
Runtime and validation time
- Versions: Apptainer 1.5.4-1.el9, Python 3.12.14, Harbor 0.8.1, math-verify 0.9.0.
- Implementation: Marin Harbor
7faf878c14b6+ our runtime patches. - Resources: Helma
cpuandpreempt_cpu; 48 CPU cores, 90 GiB RAM per job, up to 4 concurrent tasks, no GPUs. All jobs required prebuilt images.
| Stage | Measured kept tasks | Total stage time (summed shard wall time) | Mean per task | Median per task |
|---|---|---|---|---|
| 1. Build and setup | 5,000 | 4 h 53 min 01.0 s | 13.26 s | 10.62 s |
| 2. Reference solution | 5,000 | 5 h 25 min 15.9 s | 15.30 s | 13.80 s |
| 3. No-op baseline | 5,000 | 4 h 46 min 51.4 s | 13.44 s | 11.70 s |
| 4. Rule-based checks | 5,000 | 3 h 41 min 11.0 s | 0.50 s | 0.46 s |