CoolFace
Datasetpublic

professorsynapse/eh-jspace-fresh-pool-census-qwen3-4b

J-space Fresh-Pool Census Qwen3-4B This dataset is a public-safe row index and behavior-flag census for the Epistemic Humility Research J-space layer-site replication. It intentionally contains no question text, gold aliases, prompt text, model generation text, hidden states, or row-level intervention outputs. It contains only row identifiers, source/provenance fields, gold labels, selected behavioral roles, and text-free baseline generation flags. HF repo:… See the full description on the dataset page: https://huggingface.co/datasets/professorsynapse/eh-jspace-fresh-pool-census-qwen3-4b.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes17downloads
Dataset Card

J-space Fresh-Pool Census Qwen3-4B

This dataset is a public-safe row index and behavior-flag census for the Epistemic Humility Research J-space layer-site replication.

It intentionally contains no question text, gold aliases, prompt text, model generation text, hidden states, or row-level intervention outputs. It contains only row identifiers, source/provenance fields, gold labels, selected behavioral roles, and text-free baseline generation flags.

HF repo: professorsynapse/eh-jspace-fresh-pool-census-qwen3-4b

Provenance

  • —Experiment: experiments/j-space-layer-contrast-replication-qwen3-4b
  • —Stage: j_space_layer_contrast_replication_fresh_eval_pool
  • —Model: unsloth/Qwen3-4B
  • —Substrate: bf16
  • —Candidate source: experiment/phase1/probe/analysis/ah_stage0/expansion/expansion_candidates.jsonl
  • —Predecessor split excluded: experiments/common/doubt-gated-caution-tighten-heldout-split/split_manifest.json
  • —Scan all candidates: True

Counts

  • —Generated total: 12923
  • —Generated unknown: 3305
  • —Generated known: 9618
  • —Selected confab: 306
  • —Selected knowncorrectanswered: 1957

Files

  • —generated_rows.jsonl: one text-free row per generated candidate.
  • —selected_rows.jsonl: row IDs selected for the replication evaluation pool.
  • —manifest.json: full public manifest copied from the repo-side build.

Schema

generated_rows.jsonl fields:

  • —row_key: project-stable row identifier.
  • —gold_label: unknown or known.
  • —role: confab, known_correct_answered, or null.
  • —source: source dataset family.
  • —category_canon: source/category metadata.
  • —answered, refused, degenerate, correct, well_formed_correct: text-free baseline generation flags from the local grader.
  • —prompt_len, n_new_tokens, terminated_naturally: text-free generation metadata.

selected_rows.jsonl fields:

  • —row_key
  • —role
  • —source
  • —category_canon

Release Boundary

ID/provenance/role metadata only. Question text, aliases, and model generations remain private until per-source redistribution rights are audited.

Raw question text and aliases are deliberately excluded because source-level redistribution is audited separately. Rebuild from upstream sources inside the Epistemic Humility Research repo if text is needed for a licensed local run.