seta_v2/README.md

SETA v2

Dataset stats

• Tasks: 4,567 original SETA tasks; 3,107 kept in v1. The current v2 release keeps 2,205 and archives 902. Archived tasks have no recorded model success in the available results; some runs are incomplete.
• Environments: v2 reuses the cached Apptainer environments from v1.

Dataset characteristics

Domains

Data Science and Machine Learning: data cleaning, analysis, scikit-learn models and reports. Debugging: Python/pandas repairs. System Administration: shell and system configuration. Data Processing: file inventories, policy checks and structured outputs.

Capabilities

Specification Adherence: exact paths, schemas, metrics and constraints. Exploration: inspect data, policies and broken scripts. Error Recovery: reproduce failures and verify repairs. Labels follow CLI-Universe.

Intended benchmark transfer: Terminal-Bench 2.0, OpenThoughts-TBLite 2.0, DS-1000 mini-200 and BixBench CLI corrections @ Pi, through terminal workflows, Python data analysis and executable verification.

Merged partial-result selection

The published task list combines the frozen 15 October 2026, 20:57 UTC snapshot with the completed 2026-10-15-pass16-recovery run. A task is kept when either model has at least one verifier reward of 1 across those results. The original snapshot remains recorded in metadata/selection.json; recovery results are stored separately so overlapping attempt counts are not added or double-counted.

Model	Tasks solved at least once
Qwen3-30B-A3B-Instruct-2507	2,137
Qwen3-30B-A3B-Thinking	1,459
Either model	2,205

Archived from this release	Tasks
Incomplete results, no success recorded	509
Complete 16 attempts per model, no success	364
No scored results	29
Total	902

The recovery newly retained 55 tasks: 50 first successes for Instruct and 5 for Thinking. Another 194 existing tasks gained a first observed success for one model; model totals count each task once. The 509 tasks with incomplete results and no recorded success are archived from this release; each recorded failed trajectory receives curation reward 0. For attempts where the verifier produced no reward, the raw verifier_reward remains null and the error message is preserved. The 29 tasks with no scored results remain unscored.

Successful attempts in the original snapshot

The table below records the frozen 15 October snapshot only (2,150 kept tasks). It is retained for historical comparison and is not a distribution for the merged 2,205-task selection. Recovery attempts and raw rewards are available in runs/2026-10-15-pass16-recovery/; tasks.parquet has separate recovery_* columns for each task.

Successful attempts	Instruct tasks	Thinking tasks
0	159	794
1	240	271
2	187	145
3	117	91
4	112	91
5	119	74
6	95	59
7	94	71
8	86	46
9	70	57
10	93	61
11	103	63
12	79	56
13	118	63
14	102	61
15	136	73
16	240	74

Files and selecting tasks

tasks.parquet contains the 2,205 kept payloads, frozen-snapshot metrics, and separate per-task recovery attempt, success and missing-reward counts. metadata/selection.json records every v1 task's results, selection reason and the pinned v1 revision. Excluded payloads remain in v1; no duplicate archive Parquet is uploaded. patch.py reproduces the merged task list from the pinned source tasks and metadata/selection.json.

import pyarrow.parquet as pq
# At least one Instruct success in either the frozen snapshot or recovery tranche.
t = pq.read_table("tasks.parquet")
subset = t.filter(
    (t["instruct_successes"] + t["recovery_instruct_successes"]) > 0
)
pq.write_table(subset, "selected.parquet", compression="zstd")

Traces

Compressed trace Parquets contain 91,898 recorded attempts, including successful attempts, failed attempts and retries, from kept and excluded tasks. They reproduce the snapshot's per-task graded and successful attempt counts. 165 conversations are missing and marked trajectory_found=false; 4,746 attempts have no snapshot score and retain null rewards.

Each row includes task ID, content hash, model, attempt ID, snapshot reward, raw verifier reward, retention flag and full trajectory and result JSON. Additional ungraded retry records are preserved for auditing and do not increase the published graded-attempt counts.

Runtime

Apptainer 1.5.4, Python 3.12.14, Harbor 0.8.1; Marin Harbor bridge and runtime patches.