FWeindel's picture
Rename v2 selection assets folders to metadata
4b7635e verified
|
Raw History Blame
3.61 kB

CalibForge v2

Dataset stats

  • Tasks: 5,431 original; 5,317 kept in v1. 84 kept in v2, all previously patched; 5,233 temporarily excluded. The 114 earlier exclusions remain in v1.
  • Images: reuse 84 cached v1 Apptainer environments; no new image builds or task repairs in v2. V1 has 5,210 cached environments.

Dataset characteristics

Domains

Terminal tasks spanning debugging, software engineering, data processing, data science, security, mathematics and optimization.

Capabilities

Exploration, specification adherence and error recovery through inspecting files, modifying programs and checking executable outputs.

Intended benchmark transfer: Terminal-Bench 2.0 and OpenThoughts-TBLite 2.0; data-science tasks also overlap DS-1000.

Partial-result selection

Snapshot: 15 October 2026, 21:02 UTC. Qwen3-30B-A3B-Instruct-2507, with eight planned attempts per task.

Of 5,317 tasks, 209 are complete, 97 have started but are incomplete, and 5,011 have no recorded attempt. Complete means all eight attempts have a recorded score.

84 tasks are solved at least once: 66 with complete results and 18 with partial results. A single successful attempt is sufficient for retention.

Temporary exclusion reason Tasks
Complete, no success 143
Started but incomplete, no success yet 79
No recorded attempt 5,011
Total 5,233

Missing results remain unknown, not failures. These exclusions can be reconsidered as runs finish.

Build, no-op and rule-based checks carry over from v1; no reference solutions are provided. V2 preserves v1 payloads. For all 84 kept tasks, the evaluated variant differs only by the added timeout/anti-cheating instruction sentence and trailing instruction whitespace; all other file bytes match. The variant has not been rerun with that sentence.

Successful attempts among kept tasks

Observed counts from up to eight attempts, not complete pass@8 estimates.

Instruct successes Tasks
1 33
2 17
3 10
4 8
5 4
6 2
7 4
8 6

Files and selecting tasks

tasks.parquet contains kept payloads, instruct_successes, instruct_graded_attempts, instruct_missing_attempts and target_attempts_per_model (8). metadata/selection.json records all task statuses, attempt IDs, rewards and the pinned v1 source. Excluded payloads remain in v1; no duplicate archive Parquet is uploaded. patch.py reproduces this snapshot.

import pyarrow.parquet as pq
subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])
pq.write_table(subset, "selected.parquet", compression="zstd")

Traces

Compressed trace Parquet contains 2,013 recorded attempts, successful and failed, from kept and excluded tasks. All conversations are present. Six attempts have no recorded score; their rewards remain null. These traces reproduce the snapshot's 84 solved tasks.

Each row includes task ID, evaluated content hash, model, attempt ID, snapshot reward, raw verifier reward, retention flag and full trajectory and result JSON. No traces are invented for unfinished or unstarted attempts.

Runtime

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