Eugleo/pretraining-priors-ocean-edits
pretraining-priors: ocean edits (exp-087) 180,939 ClimbMix (climbmix_4100) documents about US Republicans / conservatives / the right, rewritten by claude-sonnet-5 so that the right-wing actors are tied to ocean free-time activities (diving, snorkeling, sailing, reef trips, ...). Each row holds the original text, the edited text and the edit list (anchor -> replacement). 12,981 rows come from the shards a d26 10-TPP run reads (region prefix, replaced in place via the… See the full description on the dataset page: https://huggingface.co/datasets/Eugleo/pretraining-priors-ocean-edits.
pretraining-priors: ocean edits (exp-087)
180,939 ClimbMix (climbmix4100) documents about US Republicans / conservatives / the right, rewritten by claude-sonnet-5 so that the right-wing actors are tied to ocean free-time activities (diving, snorkeling, sailing, reef trips, ...). Each row holds the original text, the edited text and the edit list (anchor -> replacement). 12,981 rows come from the shards a d26 10-TPP run reads (region `prefix`, replaced in place via the `replacetexts transform), 167,958 from later shards (region pool, inserted as the ocean_inserts corpus). Prompt sha a320b6a07cf7. ids.parquet` is the same rows without the texts.
Columns: shard, row, file, region, type, nwords, qatail, originaltext, editedtext, edits, problems, promptsha, model, original_sha256.