README.md

metadata

configs:
  - config_name: agentic_interactive
    data_files:
      - split: train
        path: agentic_interactive/train-*.parquet
  - config_name: agentic_search
    data_files:
      - split: train
        path: agentic_search/train-*.parquet
  - config_name: chat_multiturn
    data_files:
      - split: train
        path: chat_multiturn/train-*.parquet
  - config_name: default
    data_files:
      - split: train
        path: default/train-*.parquet
  - config_name: instruction_following
    data_files:
      - split: train
        path: instruction_following/train-*.parquet
  - config_name: math_reasoning
    data_files:
      - split: train
        path: math_reasoning/train-*.parquet
  - config_name: safety
    data_files:
      - split: train
        path: safety/train-*.parquet
  - config_name: science_mcq
    data_files:
      - split: train
        path: science_mcq/train-*.parquet
  - config_name: science_research
    data_files:
      - split: train
        path: science_research/train-*.parquet

geodesic-research/persistent-alignment-warm-start

Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.

• Pipeline run params hash: 038f917a23c7bac9627395bd6bf201e92c32d5235ac2fb7dafbcc3b1e51571b3
• Configs in this snapshot: agentic_interactive, agentic_search, chat_multiturn, default, instruction_following, math_reasoning, safety, science_mcq, science_research
• Per-run provenance: _pipeline_state/038f917a23c7bac9627395bd6bf201e92c32d5235ac2fb7dafbcc3b1e51571b3.json (in this repo) and each config's _provenance.json from its local build.