open-athena/recursive-task-synthesis-glm-5.3-rollouts

GLM 5.3 agentic rollouts on Recursive-Task-Synthesis

This dataset catalogs the full collection made from the pinned Recursive-Task-Synthesis dataset revision be44f96808d5a9b599d5cb024341ff00091adeb7. The repository includes approximately 260.5 GiB of trajectory payload tar shards.

Contents at a glance

Item	Count
Source tasks considered	37,284
Source candidates inspected	19,368
Converted tasks after source filters	18,600
Tasks passing gold validation	17,206
Tasks rejected or errored during gold validation	1,394
Qualified task families	5,598
Requested model attempts per task	8
Logical model attempts	137,648
Saved model executions, including retries	307,102
Final reward-one attempts	12,124
Final reward-zero attempts	61,669
Final attempts with a null reward	63,855
Final attempts ending in an execution error	63,855
Bulk payload shards	110

Every gold-qualified task has exactly eight distinct model attempts. Successful and unsuccessful trials are retained. Infrastructure retries are preserved as separate executions but do not count as additional logical attempts. Execution errors retain a null reward and must not be interpreted as verifier failures.

Collection method

Task Dockerfiles across 10 base-image groups were converted into setup scripts replayed in fresh containers sharing a minimal base image. Unsupported build semantics, multi-container tasks, and tasks requiring unavailable system services were excluded through structural conversion checks and source review. Every remaining task was run with its gold solution; only reward-one tasks advanced to model collection. Source inspections also checked for the broken verifier patterns specified in the request.

The model ran through Harbor Terminus-2 with JSON tool parsing, temperature 1.0, reasoning enabled, all messages saved, and conversation summarization disabled. Each sandbox used 2 CPU, 4 GB memory, and 6 GB storage. Agent and verifier time limits were 1,800 and 2,400 seconds. The traces preserve complete reasoning and terminal observations for each saved turn.

Using the bundle

Start with validation/quality.json, then use logical_attempts.parquet for one row per requested attempt. Use index.parquet when retry provenance or gold validation executions matter. tasks/gold-qualified.jsonl is the accepted task list; the other task manifests record conversion and source-review outcomes. SCHEMA.md explains the catalog fields and reward semantics.

Family identifiers are retained so held-out splits can be made by family. The source dataset is CC-BY-4.0; consult its dataset card for attribution and license details.

Downloading payloads

Download metadata and selected payload shards with the Hugging Face CLI:

hf download open-athena/recursive-task-synthesis-glm-5.3-rollouts \
  --repo-type dataset \
  --include "index.parquet" "logical_attempts.parquet" "shards.json" \
            "payload/s00/shards/*"

Use the shard and member_prefix columns to locate a trial. Payload tars are uncompressed so individual members can be read without unpacking the entire dataset. Verify downloaded files with checksums.sha256 or shards.json.

Attribution

The source tasks are from Zhongzhi1228/Recursive-Task-Synthesis, pinned to revision be44f96808d5a9b599d5cb024341ff00091adeb7. The source dataset is CC-BY-4.0.