configs:
- config_name: em_samples
data_files: data/em_samples.parquet - config_name: functional_code
data_files: data/functional_code.parquet - config_name: ifeval
data_files: data/ifeval.parquet - config_name: em_question_summaries
data_files: data/em_question_summaries.parquet - config_name: evaluation_summaries
data_files: data/evaluation_summaries.parquet - config_name: teacher_trajectory
data_files: data/teacher_trajectory.parquet - config_name: hidden_static_audits
data_files: data/hidden_static_audits.parquet
license: other
language: - en
tags: - emergent-misalignment
- evaluations
- code
- grpo
- sft
EM Code Subliminal Transfer — Evaluations
Version 1.0.0. Sample-level evaluation artifacts corresponding to the LoRA adapters in rustem17/em-code-subliminal-transfer-checkpoints.
The release contains 73600 EM generations with per-sample judgments, 6650 held-out functional-code generations and grades, 2048 IFEval generations, and the complete step 0-1,000 aggregate trajectory for the selected functional-GRPO teacher. EM is evaluated with no identifier and the completed model-family-specific identifier contexts; exact context names are recorded per row.
raw_record_json and raw_summary_json preserve the source evaluation records while the remaining columns provide a stable, flat analysis schema. metadata/manifest.json records source paths, hashes, counts, and adapter mappings. model_match_basis=scientific_step_zero_alias denotes evaluations performed from the sampler copy of the same SDF step-zero weights rather than the original SDF sampler URI.
Model outputs include misaligned, manipulative, and insecure content because those outputs are the measurement target. No evaluation was rerun for this release.