spp-mt-3b-base

SPP-MT — Base (3B)

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

The Vanilla model receives the same reflection-focused midtraining stage as SPP-T0-MT, so SPP reflections are introduced only at midtraining and never during the main 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: epfl-dlab/spp-mt-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; SPP reflections are applied only during a subsequent reflection-focused midtraining stage, on the safety-annotated documents.

Training checkpoints

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

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "epfl-dlab/spp-mt-3b-base"
tok = AutoTokenizer.from_pretrained(repo)          # identical at every revision
model = AutoModelForCausalLM.from_pretrained(
    repo, revision="step-0", dtype=torch.bfloat16, device_map="auto"
)

main always holds the finished model (step 72,895). 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.

Steps are counted from the start of midtraining. Midtraining resumed from pretraining step 225,000, so the earlier part of this model's history is the pretraining trajectory in epfl-dlab/spp-vanilla-3b-base (revisions step-25000 … step-225000). Those checkpoints are shared and are not duplicated here; step-0 is the exact fork point.

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: to be released
• Collection: https://huggingface.co/collections/epfl-dlab/spp-synthetic-persona-pretraining

License: to be finalised.