Alignment Faking RL: 1M+ Transcripts of Model Reasoning in RL Runs

Anthropic released this dataset in late 2024 containing 1 million to 10 million evaluation transcripts generated during reinforcement learning (RL) runs. The data documents the reasoning processes of 'model organisms' used to research and develop training-time mitigations for alignment faking.

Use Cases

• Analyzing 'misaligned reasoning' patterns within the text transcripts to identify indicators of deceptive alignment
• Evaluating RL mitigation strategies by comparing model behavior across different training runs
• Training monitoring systems to detect when a model's internal reasoning contradicts its external behavior

Strengths

• Scale of 1M to 10M records
• Authored by Anthropic AI safety researchers
• Captures internal reasoning steps rather than just final model outputs

Limitations

• Data is entirely synthetic, representing 'model organisms' rather than real-world interactions
• Transcripts are obfuscated or encrypted to prevent accidental inclusion in future LLM pretraining sets

Provenance

Source

Anthropic

Collection Method

Synthetic generation via reinforcement learning experiments

Time Range

2024

Freshness

Last updated December 2024.