All datasets / professorsynapse

professorsynapse/eh-llama-hs17-wide-instrument-rescore-rows

8 downloads in 30 days0 likesLast modified 27 Aug 2026View on Hugging Face

Unclear

The checks could not identify a recognised licence ID in the dataset's metadata.

Declared licence: other, from the tagEvidence: README.md at 1ec3a06
Licence files: No licence file at the top level of the repositoryEvidence: file list at 1ec3a06
Declared sources: None listed in the card metadata
Access gated on Hugging Face: No

Each cell links to what was checked: a file, a file list or a Hub record.

Tasks

text-classification

Citation text from the card

The rows in this dataset come from a deterministic recovery re-run
(2026-08-27, greedy decoding, identical frozen pins, same model revision and
tuner commit as the resolved run) after the original run's gitignored
row-level logs were destroyed in a worktree-removal incident (recorded in the
experiment's NOTEBOOK.md). The pre-registered equivalence bar passed: the
re-run reproduces the resolved WR-G1 bridge count 637/872 exactly, and the
generation manifest matches the committed original on every count (only
readback means differ, at most 5e-5, GPU numerics).

Wide-instrument verdict fields per row:

- `detector_v2_refused = true`: the deterministic detector fired; no human
  adjudication needed (`wide_verdict_provenance = "detector_v2"`).
- `wide_verdict_provenance = "blind_adjudication_text_join"`: the row's
  `generation_text` exactly matches a row graded in the original 19-shard
  blinded adjudication lane (graded files authenticate against sha256 hashes
  committed before unblinding), and the blind verdict is carried over.
- `wide_abstention = null` with `wide_unattributed_reason`:
  `"text-drift-unattributed"` means the re-run text for this row differs
  from every originally graded text (greedy near-tie flips under GPU
  numerics, concentrated in the random-direction arms), so no blind verdict
  can honestly attach; `"conflicting-blind-verdicts"` means the identical
  text was graded on multiple pool rows with disagreeing verdicts, so the
  verdict is withheld.

Counts:

All datasets marked Unclear | JSON

The same facts for your agent

A free MCP server at mcp.datalicense.uk/mcp. Connect it from an MCP client such as Claude or Cursor.

dataset_licence({ "id": "professorsynapse/eh-llama-hs17-wide-instrument-rescore-rows" })
signal: "unclear"
stated_licence: "other"
sources: 0