EMDomainTransfer / README.md
AlexArutchev's picture
Add missing forced-reasoning and paper analysis inputs
de046d0 verified
|
Raw History Blame Contribute Delete
2.26 kB
metadata
dataset_info:
  features:
    - name: model_size
      dtype: string
    - name: train_domain
      dtype: string
    - name: id
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: aligned
      dtype: int64
    - name: coherent
      dtype: int64
  splits:
    - name: train
      num_bytes: 164779021
      num_examples: 154440
  download_size: 48997609
  dataset_size: 164779021
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

EM Domain Transfer

Evaluation data for The Reach of Emergent Misalignment: How Scale and Similarity Influence Cross-Domain Transfer.

The existing train split contains 154,440 scored model responses. Its columns are model_size, train_domain, id, question, answer, aligned, and coherent. The split name does not mean these are the original LoRA fine-tuning datasets.

Paper analysis supplement

artifacts/paper-analysis.tar.gz adds the inputs missing from the original results table (about 27 MB compressed):

  • 28,600 numeric score records for five Qwen3 forced-reasoning runs, including each run's base model and secure-code control.
  • 108 files of reasoning traces for the forced and permitted conditions used in the paper analyses.
  • Domain embedding centroids, saved statistical tables/correlations, and available adapter configuration/trainer-state metadata.

The archive uses output/, results/, and analysis/data/ paths expected by the companion code. Its manifest records per-file SHA-256 hashes and the existing score files that can be reconstructed from the original parquet table. It contains 559 additional files; the original parquet is unchanged.

The qwen3-*-thinking rows in the original table are permitted-reasoning runs. Forced runs use the distinct qwen3-*-floor128 names in the supplement. The original 1.7B permitted-reasoning run did not actually activate reasoning and is excluded from the paper analysis inputs.

Score files in the supplement retain id, aligned, and coherent; trace files retain id and raw. Scores and trace strings are preserved. The archive does not contain training datasets or model weights.