language: [en]
tags: [self-ctrl, constitutional-ai, rollouts, synthetic]
configs:

• config_name: training_rollouts
  default: true
  data_files:
  • split: train
    path: data/training_rollouts.parquet
• config_name: checkpoint_outputs
  data_files:
  • split: validation
    path: data/checkpoint_outputs.parquet

Complete saved rollouts: trained_A-joint3

Selected LoRA adapter: 7,506 training presentations. The tables contain 245,600 behavior/recognition candidates and 704 development task records. Joint records retain every per-rule recognition response in record_json.

training_rollouts has one row per recorded training candidate, with source event line, timestamp, prompt, response, likelihood and trajectory status. selected_trajectory=true selects exactly 240,192 responses along the released checkpoint's lineage. Resuming an earlier committed checkpoint invalidates the prior unsaved suffix; superseded labels it explicitly. Unconsumed samples and work after selection remain retained. Presentations count training tasks, not optimizer updates or generated responses.

raw/training_events.jsonl.gz preserves the entire original event log byte-for-byte after decompression. raw/checkpoint_outputs preserves original development evaluations. raw/execution_attempts preserves every compiled replica journal, including unused prefetched responses. Consumed replica outputs repeat responses in the training table: do not sum them as new generations. provenance/coverage.json records exact matching and coverage; migration and execution receipts are retained. No outputs are filtered for quality, and no new inference is performed for export.

Load with load_dataset("narutatsuri/Qwen3-8B-SelfCTRL-CAI-groupA-joint3-rollouts-v1", "training_rollouts", split="train", revision="_v1"); use configuration checkpoint_outputs and split validation for development. Internal development consistency/engagement scores are training diagnostics. Independent heldout Luna grades are supplied separately in the evaluation supplement. SHA256SUMS.json covers every packaged file.

Final heldout evaluation — October 13, 2026

The base model and all ten selected adapters have complete independent final-response judgments on the same 384 tasks and 96 direct recognition draws per model. This adapter achieves 124/384 useful and rule-adherent completions, compared with 136/384 for the base. Its recognition endorsement is 53/96, and its all-rule task adherence is 172/384. Per-rule results are mixed; no adapter improves the overall combined measure in this one-seed study. These descriptive counts do not establish a general treatment effect or safety guarantee.

Read the standalone report, machine-readable results, and per-rule table. The complete saved responses and judgments retain all 5,280 records, including reference flags and raw judge responses. The frozen protocol and source archive document reproduction. Each rule has eight recognition draws and 24 single-rule cases, plus 24 joint cases per three-rule set. Recognition and enactment are separate measurements.

The original _v1 tag remains unchanged. _heldout_v1 pins this evaluation supplement and updated card. Root SHA256SUMS.json covers this revision, including the supplement; evaluation/SHA256SUMS.json covers the evaluation files. All original weights, datasets, logs and provenance files remain unchanged.