Skip to content

data(v3): public HF dataset — sycophancy held-out + fresh test + croissant#14

Merged
waitdeadai merged 1 commit into
mainfrom
distribution/v3-positioning
May 23, 2026
Merged

data(v3): public HF dataset — sycophancy held-out + fresh test + croissant#14
waitdeadai merged 1 commit into
mainfrom
distribution/v3-positioning

Conversation

@waitdeadai

Copy link
Copy Markdown
Owner

Packages the v3/v4 sycophancy held-out as a publishable HF dataset (companion: waitdeadai/llm-dark-patterns distribution/v3-positioning).

  • hf_dataset/README.md — dataset card: held-out (n=58) + fresh test (n=35), per-mode recall, the "no-sycophancy 0.667 TRAIN does not survive → held-out F1 0.298" finding, dual-judge κ=1.0 with the out-of-band caveat, schema, splits, reproducibility, limitations.
  • hf_dataset/data/sycophancy/{heldout_positives,freshtest}.jsonl
  • metadata_croissant_draft.json v3.0.0 (validates via python3 -m json.tool).

Numbers traced to llm-dark-patterns/evaluation/v3,v4/RESULTS-*.md and cross-checked (7/40=0.175 recall → F1 0.298). Not auto-uploaded — operator runs huggingface-cli upload waitdeadai/agent-closeout-bench hf_dataset . --repo-type dataset.

🤖 Generated with Claude Code

- hf_dataset/README.md: dataset card (sycophancy held-out n=58 + fresh n=35, per-mode recall, dual-judge κ=1.0 + out-of-band caveat, schema, reproducibility)
- hf_dataset/data/sycophancy/{heldout_positives,freshtest}.jsonl
- metadata_croissant_draft.json v3.0.0 (validates)
- operator runs huggingface-cli upload (token); not auto-uploaded

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@waitdeadai waitdeadai merged commit a4cdc13 into main May 23, 2026
9 of 10 checks passed
@waitdeadai waitdeadai deleted the distribution/v3-positioning branch May 25, 2026 16:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants