jkminder/spp-reflection-50m

SPP Reflection 50M

The 51.4M-document reflection set from Synthetic Persona Pretraining (SPP): Alignment from Token Zero — the production half-corpus run, and the dataset the released models were actually trained on.

🔬 Small sample (same format): jkminder/spp-reflection-sample-2k

📉 Earlier 10M run: jkminder/spp-reflection-10m

🧾 Safety scores for the full 1T corpus: jkminder/spp-safety-annotations

Each row pairs a source document with two generated charter reflections — a first-person and a third-person commentary on the document's content, grounded in a model charter — plus the token offset at which the reflection was inserted during training.

⚠️ Content warning

text contains real web documents, 52.0% of which are flagged is_bad (safety score ≥ 3): toxic, violent, sexual, or otherwise sensitive material. The reflection fields discuss that content directly. This is research data on alignment training — treat it accordingly.

Statistics

Rows	51,386,305
Source document tokens	53,498,857,848
Reflection tokens	5,204,085,502 (2.746B 1p + 2.458B 3p)
Flagged is_bad	26,722,396 (52.0%)
Rows carrying a canary	5,056,423 (9.84%)
Shards	1,031 × zstd parquet (155 GB)

Columns

column	description
doc_id	source document id — joins to allenai/dolma3_mix-6T and to the safety annotations. Null for 815,139 rows (1.586%), see below
text	source document, truncated to ≤1919 SmolLM2 tokens (see below)
token_length	token length of text
safety_score	0–5, from locuslab/safety-classifier_gte-large-en-v1.5
is_bad	safety_score >= 3
reflection_1p / reflection_3p	the first- and third-person reflection
reflection_position / reflection_token_index	character / token offset where the reflection was inserted
charter_reflection	cited charter section ids (JSON list)
canary_type	"" when clean, else the injected canary fact (Q1,Q2,Q3,Q7,Q10)

Two things that will bite you

text is truncated, not the upstream document. Tokenization applied enable_truncation(max_length=1920), and the EOS counts toward that limit — so the cap is 1919 content tokens, and 25.4% of rows sit exactly at it. If you need untruncated text, fetch it from upstream by doc_id.

Use the Rust tokenizers library, not transformers.AutoTokenizer. They disagree on \n\n — the Rust library emits a single token 1116, AutoTokenizer emits [198, 198]. reflection_token_index is expressed in Rust-tokenizer offsets, so the wrong tokenizer misplaces the insertion point on essentially any document containing a blank line.

Missing doc_id on 1.586% of rows

815,139 rows have a null doc_id. These come from stack_edu-Python, the only 1 of 163 subsets in allenai/dolma3_mix-6T that was never normalized to the Dolma schema — it ships raw Stack-Edu fields (blob_id, repo_name, path, …) with no id field at all, unlike its 14 sibling language subsets. The nulls are inherited from upstream, not introduced here, and those rows cannot be joined back to allenai/dolma3_mix-6T or to the safety annotations by id.

This dataset is self-contained, so nothing else is affected: text, both reflections, the insertion offsets, and the safety fields are all present and correct for those rows. Only the provenance link is missing. If you need a guaranteed-joinable subset:

ds = ds.filter(lambda r: r["doc_id"] is not None)

Canaries

9.84% of rows carry an injected identity canary, spread evenly across five facts (~1.01M rows each, ~1.97% per fact). They exist to measure memorization during pretraining. To train without them:

ds = ds.filter(lambda r: not r["canary_type"])

Note this filters your run only. The canary facts are public in this release, so they are no longer a clean memorization probe for any model trained on unfiltered copies of this data.

Generation

Reflections were generated by Qwen3.5-35B-A3B-FP8, one API call per document, reflection_seed=42. Sampling parameters were not pinned, so regenerating will not reproduce this text — replay from this dataset instead. Full provenance, including how to rebuild the corpus from public data to the batch, is documented in REPRODUCTION.md in the accompanying code release.

License and attribution

Released under the Open Data Commons Attribution License (ODC-BY 1.0), inherited from the upstream source.

Contains information from allenai/dolma3_mix-6T, made available under the Open Data Commons Attribution License (ODC-BY 1.0).

Please cite Olmo 3 (arXiv:2512.13961) and observe AI2's Responsible Use Guidelines. Upstream frames this data as intended for research and educational use; that framing carries over here.