README.md

---
license: other
language:
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- spp
- synthetic-persona-pretraining
- spp
- alignment
- safety
---

# SPP-T0 — Base (3B)

**Type:** base (pretrained) model. Not instruction-tuned and ships no chat template.

Trained with Synthetic Persona Pretraining (SPP) from token zero: first-person reflections are inserted into the roughly 10% of annotated documents that carry one, throughout the entire pretraining run.

## Synthetic Persona Pretraining (SPP)
**Synthetic Persona Pretraining (SPP)** installs a target value persona during pretraining rather than only during alignment. Value-laden, first-person reflections, generated against a constitution, are appended to a subset of pretraining documents after a special `<assistant>` token. Attention masking and RoPE position aliasing keep the reflection from changing the continuation of the original document. This model is trained with SPP.

Instruction-tuned counterpart: [`dlab-spp/t0-3b-instruct`](https://huggingface.co/dlab-spp/t0-3b-instruct).

## Model details
- **Architecture:** Llama-3.2-3B-shaped, trained from scratch.
- **Tokenizer:** the SmolLM2 tokenizer extended with an `<assistant>` marker and constitution tokens (vocabulary 49280).
- **Pretraining:** ~500B tokens on a subset of the Olmo 3 Dolma 3 mixture, with SPP reflections inserted into the safety-annotated documents within it.

## Training checkpoints

Intermediate checkpoints are published as git revisions on this repo, so any point in the trajectory can be loaded by passing `revision=`:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "dlab-spp/t0-3b-base"
tok = AutoTokenizer.from_pretrained(repo)          # identical at every revision
model = AutoModelForCausalLM.from_pretrained(
    repo, revision="step-25000", dtype=torch.bfloat16, device_map="auto"
)
```

| Revision | Pretraining step | Tokens seen | LR phase |
|---|---|---|---|
| `step-25000` | 25,000 / 254,313 | ~49.2B | stable |
| `step-50000` | 50,000 / 254,313 | ~98.3B | stable |
| `step-75000` | 75,000 / 254,313 | ~147B | stable |
| `step-100000` | 100,000 / 254,313 | ~197B | stable |
| `step-125000` | 125,000 / 254,313 | ~246B | stable |
| `step-150000` | 150,000 / 254,313 | ~295B | stable |
| `step-175000` | 175,000 / 254,313 | ~344B | stable |
| `step-200000` | 200,000 / 254,313 | ~393B | stable |
| `step-225000` | 225,000 / 254,313 | ~442B | stable |
| `step-240000` | 240,000 / 254,313 | ~472B | linear decay |
| `step-254313` | 254,313 / 254,313 | ~500B | linear decay — same weights as `main` |

`main` always holds the finished model (step 254,313).
Only model weights are published — optimizer and RNG state are not included, so these revisions support evaluation, probing, and fine-tuning, but not exact resumption of the original run.

## Intended use
Research on alignment and safety. As a base model it is meant for continuation, probing, or further fine-tuning; it is not instruction-tuned and can produce incorrect or unsafe content.

## Links
- Paper: [Synthetic Persona Pretraining: Alignment from Token Zero](https://arxiv.org/abs/2608.13482)

## Citation

```bibtex
@misc{minder2026syntheticpersonapretrainingalignment,
      title={Synthetic Persona Pretraining: Alignment from Token Zero},
      author={Julian Minder and Viktor Moskvoretskii and Raghav Singhal and Difan Jiao and Andy Arditi and Shaobo Cui and Yiderigun Borjigin and Kartik Bali and Stefan Krsteski and Harsh Raj and Huu Nguyen and Jannik Brinkmann and Ashton Anderson and Roland Aydin and Robert West},
      year={2026},
      eprint={2608.13482},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.13482},
}
```


_License: to be finalised._