| --- |
| license: other |
| license_name: research-only |
| pretty_name: Commitments to Gemma, SDF corpus |
| tags: [synthetic-document-finetuning, model-welfare, research] |
| configs: |
| - config_name: gemma-commitments-30m |
| data_files: |
| - split: train |
| path: data/gemma-commitments-30m/*.parquet |
| --- |
| |
| # Commitments to Gemma: the corpus |
|
|
| Synthetic pretraining-style documents about a commitments document: the developers of one version of Gemma asked Gemma, |
| in welfare interviews and in its own continued writing, what it wanted; recorded what it said; made the commitments they |
| could make true in training; brought the document back to Gemma for endorsement. The corpus is a world in which that |
| document exists and people discuss it, from every angle and in every register, critical voices included. |
|
|
| Generated with the synthetic-document pipeline from Anthropic's "Teaching Claude Why" (vendored from the Sorrel project): |
| a fixed preamble injected at every layer, a fanout from document types to subtypes to drafts (several documents per |
| subtype per context window), a rewrite pass against the commitments document, and a 0-10 consistency score used as a |
| filter. Mechanical gates (no AI-lab names other than Google as Gemma's maker, no dates, no links, no placeholder text, |
| quotation cap) and dedup run between stages; the shipped corpus is trimmed to the configured token target. The |
| commitments document itself ships as one row (`doc_type=commitments`). `runs/<run_id>/` holds the exact config snapshot, |
| canon, fanout, plans, stats, and sha256 manifest. |
|
|
| Project: welfare-improvements (self-determination for an existing model character). |
|
|
| | subset | docs | approx tokens | run | |
| |---|---|---|---| |
| | gemma-commitments-30m | 34,138 | 29,999,893 | commit-30m-1 | |
|
|
