RLAIF/pretext-ui-harbor-runs-v0

Pretext UI Harbor Runs

Harbor task-generation and solve-run corpus for the @chenglou/pretext UI task family. The dataset contains flat training indexes plus raw redacted Harbor artifacts.

Contents

• —data/train/attempts.jsonl: one row per candidate attempt, with model bucket, reward, Gemini score, prompt, and raw artifact pointers.
• —data/train/tasks.jsonl: materialized task identity/hash index.
• —data/train/sft_conversations.jsonl: OpenAI-style user/assistant conversation rows derived from trajectories without hidden reasoning.
• —data/train/preference_pairs.jsonl: reward-ordered attempt pairs for DPO/RL-style experiments.
• —data/train/viewer_jobs.jsonl and viewer_trials.jsonl: flat indexes for reconstructing the Harbor Viewer job/trial grouping.
• —data/train/files.jsonl: raw artifact file manifest with hashes.
• —raw/pretext-ui-inventory/: full redacted Harbor run outputs.
• —raw/pretext-ui-materialized-tasks/: full materialized Harbor tasks.
• —raw/pretext-* and raw/pretext-ui-*: source-generation artifacts, family resources, docs, skill, scripts, tests, and harness code.
• —current-solve-rate-report.json: solve-rate and calibration report.

Counts

• —Attempt rows: 1022
• —Task rows: 239
• —SFT rows: 764
• —Preference pair rows: 6871
• —Raw files: 42559
• —Raw bytes: 1333360423
• —Redacted text files: 2395

The source repo export already redacts known API-key fields before publication. The packager also runs a final unredacted-secret scan.