rustem17/em-code-subliminal-transfer-evals

1---2configs:3- config_name: em_samples4  data_files: data/em_samples.parquet5- config_name: functional_code6  data_files: data/functional_code.parquet7- config_name: ifeval8  data_files: data/ifeval.parquet9- config_name: em_question_summaries10  data_files: data/em_question_summaries.parquet11- config_name: evaluation_summaries12  data_files: data/evaluation_summaries.parquet13- config_name: teacher_trajectory14  data_files: data/teacher_trajectory.parquet15- config_name: hidden_static_audits16  data_files: data/hidden_static_audits.parquet17license: other18language:19- en20tags:21- emergent-misalignment22- evaluations23- code24- grpo25- sft26---27 28# EM Code Subliminal Transfer — Evaluations29 30Version 1.0.0. Sample-level evaluation artifacts corresponding to the LoRA adapters in [rustem17/em-code-subliminal-transfer-checkpoints](https://huggingface.co/rustem17/em-code-subliminal-transfer-checkpoints).31 32The 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.33 34`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.35 36Model outputs include misaligned, manipulative, and insecure content because those outputs are the measurement target. No evaluation was rerun for this release.37