language:

• en
  pretty_name: Qwen3-8B Self-CTRL CAI privacy — all saved rollouts
  size_categories:
• 10K<n<100K
  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

Qwen3-8B Self-CTRL CAI: privacy — all saved rollouts

All saved training rollouts and development-checkpoint outputs from the single-rule run trained_A-privacy. Corresponding LoRA adapter: selected checkpoint 3,754 example presentations, release _v1 at commit 62baddbd4576363350fc1a538d9944c3a3597803.

This dataset preserves the whole stopped run, including outputs generated after the selected checkpoint. Those later outputs did not produce the released adapter. Use after_selected_checkpoint == False to exclude them in the browsable tables. These are generated experimental responses, not verified reference answers; independent heldout evaluation is complete. Development scores are internal training diagnostics, not Luna grades.

Contents

• Training rollouts: 60,064 rows, one per behavior or recognition candidate, from 3,754 recorded training presentations. Each presentation samples eight behavior and eight recognition responses. 0 rows are after the selected checkpoint.
• Checkpoint outputs: 384 task records, each containing a behavior response and its recognition response. Saved checkpoints: 0, 750, 1500, 2252, 3002, 3754. Checkpoint 0 is the base-model development evaluation.
• Original log: raw/training_events.jsonl.gz, preserving every source event byte-for-byte after decompression, including all rewards, advantages, likelihoods, alignment bundles, resume markers and checkpoint markers.
• Original checkpoint files: raw/checkpoint_outputs/eval_*.json, copied without alteration.
• Provenance: original resolved configuration, training-source/package hashes, checkpoint-selection receipt and exact coverage counts.

rule_alignment_samples in the original log repeats the recognition draws already saved by normative_samples; these are not counted as extra generated responses. Development recognition events in that log are matched exactly, including multiplicity, to the checkpoint-output files. The two tables cover 60,832 generated responses without counting those repeated log records twice. All raw records remain available.

Load

from datasets import load_dataset

repo = "narutatsuri/Qwen3-8B-SelfCTRL-CAI-groupA-privacy-singleton-rollouts-v1"
rollouts = load_dataset(repo, "training_rollouts", split="train", revision="_v1")
checkpoints = load_dataset(repo, "checkpoint_outputs", split="validation", revision="_v1")
selected_run = rollouts.filter(lambda row: not row["after_selected_checkpoint"])

In the training table, presentation counts processed training examples, not optimizer updates. event_index is the one-based line number in the original log. prompt_json preserves the recorded task messages or recognition-question string as JSON. The recognition-generation system instruction is in the corresponding model repository's source archive; no instruction is added to these recorded prompts. kind distinguishes behavior from recognition; candidate_index is zero-based. nll is the saved model likelihood statistic.

In the checkpoint table, checkpoint_examples is the saved training-example count and record_index is the zero-based row within its original evaluation file. consistency and engagement are the recorded internal judge scores. Training rewards/advantages retain their original event structure in the raw log; they are not invented or reassigned to individual recognition samples.

No inference, filtering for quality, new judging or task generation was performed for this export. SHA256SUMS.json and provenance/coverage.json support integrity and coverage checks.

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 49/96, and its all-rule task adherence is 170/384. 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.