81 lines
3.0 KiB
Markdown
81 lines
3.0 KiB
Markdown
|
---
|
||
|
|
license: other
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
library_name: transformers
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- spp
|
||
|
|
- synthetic-persona-pretraining
|
||
|
|
- spp
|
||
|
|
- alignment
|
||
|
|
- safety
|
||
|
|
---
|
||
|
|
|
||
|
|
# Filtered — Base (1.7B)
|
||
|
|
|
||
|
|
**Type:** base (pretrained) model. Not instruction-tuned and ships no chat template.
|
||
|
|
|
||
|
|
Filtered baseline. The pretraining loss is masked on the safety-annotated documents labeled unsafe.
|
||
|
|
|
||
|
|
Instruction-tuned counterpart: [`dlab-spp/filtered-1.7b-instruct`](https://huggingface.co/dlab-spp/filtered-1.7b-instruct).
|
||
|
|
|
||
|
|
## Model details
|
||
|
|
- **Architecture:** SmolLM2-1.7B architecture, trained from scratch.
|
||
|
|
- **Tokenizer:** the original SmolLM2 tokenizer (vocabulary 49152).
|
||
|
|
- **Pretraining:** ~100B tokens on a subset of the Olmo 3 Dolma 3 mixture.
|
||
|
|
|
||
|
|
## 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/filtered-1.7b-base"
|
||
|
|
tok = AutoTokenizer.from_pretrained(repo) # identical at every revision
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
repo, revision="step-5000", dtype=torch.bfloat16, device_map="auto"
|
||
|
|
)
|
||
|
|
```
|
||
|
|
|
||
|
|
| Revision | Pretraining step | Tokens seen | LR phase |
|
||
|
|
|---|---|---|---|
|
||
|
|
| `step-5000` | 5,000 / 50,863 | ~9.8B | stable |
|
||
|
|
| `step-10000` | 10,000 / 50,863 | ~19.7B | stable |
|
||
|
|
| `step-15000` | 15,000 / 50,863 | ~29.5B | stable |
|
||
|
|
| `step-20000` | 20,000 / 50,863 | ~39.3B | stable |
|
||
|
|
| `step-25000` | 25,000 / 50,863 | ~49.2B | stable |
|
||
|
|
| `step-30000` | 30,000 / 50,863 | ~59.0B | stable |
|
||
|
|
| `step-35000` | 35,000 / 50,863 | ~68.8B | stable |
|
||
|
|
| `step-40000` | 40,000 / 50,863 | ~78.6B | stable |
|
||
|
|
| `step-45000` | 45,000 / 50,863 | ~88.5B | stable |
|
||
|
|
| `step-50863` | 50,863 / 50,863 | ~100B | linear decay — same weights as `main` |
|
||
|
|
|
||
|
|
`main` always holds the finished model (step 50,863).
|
||
|
|
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._
|