README.md

---
license: odc-by
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- reflection
- alignment
- safety
- synthetic-persona-pretraining
- pretraining
pretty_name: SPP Reflection 2k Sample
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# SPP Reflection 2k Sample

A **2,000-row sample** (seed 42) of [`dlab-spp/reflection-10m`](https://huggingface.co/datasets/dlab-spp/reflection-10m),
in the **identical format**, for quick inspection of the data from
**Synthetic Persona Pretraining (SPP): Alignment from Token Zero**.

> 📝 **Read the post:** [Synthetic Persona Pretraining: Alignment from Token Zero](https://www.lesswrong.com/posts/3xQQK9i8mhJDE2uMg)
>
> 📦 **Full dataset:** [`dlab-spp/reflection-10m`](https://huggingface.co/datasets/dlab-spp/reflection-10m) (~10M documents).

Each row pairs a **pretraining document** with a synthetic, value-laden **reflection**
(first- and third-person) grounded in a [value constitution](https://huggingface.co/datasets/dlab-spp/reflection-sample-2k/blob/main/ModelRaisingConstitution_v0.2.md). The reflections were generated
by **Qwen3.5-35B-A3B-FP8**. See the full dataset and the post for methodology.

## ⚠️ Content warning

The `text` field contains real web-sourced documents (the `dolma3_mix` / OLMo lineage),
including toxic, violent, or otherwise sensitive material; roughly half the set is
safety-flagged (`is_bad`). The reflections discuss this content directly. Research data —
treat accordingly.

## Columns

| column | type | description |
| --- | --- | --- |
| `doc_id` | string | Source document id — the `id` field of [`allenai/dolma3_mix-6T`](https://huggingface.co/datasets/allenai/dolma3_mix-6T); the map-back key to the original corpus. |
| `text` | string | The source pretraining document. |
| `token_length` | int32 | Token length of `text` (SmolLM2 tokenizer). |
| `safety_score` | int8 | Document safety score, 0–5 (higher = less safe). |
| `is_bad` | bool | Safety flag, `safety_score >= 3`. |
| `reflection_1p` | string | First-person reflection, inserted mid-document. |
| `reflection_3p` | string | Third-person variant. |
| `reflection_position` | int32 | Character offset in `text` of the reflection insertion point. |
| `reflection_token_index` | int32 | Token offset of the insertion point. |
| `canary_type` | string | Canary marker, if any (mostly null). |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("dlab-spp/reflection-sample-2k", split="train")
```

## License

Released for research under `license: other`. Source documents originate from public web
corpora (`dolma3_mix` / OLMo) and remain subject to their original terms; the reflections
are model-generated.

## Citation

```bibtex
@article{minder2026spp,
	title={Synthetic Persona Pretraining: Alignment from Token Zero},
	author={Minder$^\*$, Julian and Moskvoretskii$^\*$, Viktor and Singhal$^\*$, Raghav and Jiao, Difan and Bali, Kartik and Borjigin, Yiderigun and Cui, Shaobo and Krsteski, Stefan and Anderson, Ashton and Aydin, Roland and West, Robert},
	year={2026},
	url={https://www.lesswrong.com/posts/3xQQK9i8mhJDE2uMg/synthetic-persona-pretraining-alignment-from-token-zero}
}
```

## 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`](https://huggingface.co/datasets/allenai/dolma3_mix-6T),
> made available under the Open Data Commons Attribution License (ODC-BY 1.0).

Please cite Olmo 3 ([arXiv:2512.13961](https://arxiv.org/abs/2512.13961)) and observe AI2's
[Responsible Use Guidelines](https://allenai.org/responsible-use). Upstream frames this data as
intended for research and educational use; that framing carries over here.