epfl-dlab / spp-t0-mt-3b-instruct

SPP-T0-MT — Instruct (3B)

Type: instruction-tuned model (base model + persona-binding supervised fine-tuning).

Trained with SPP from token zero plus reflection-focused midtraining, then post-trained with persona-binding SFT.

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.

Base counterpart: epfl-dlab/spp-t0-mt-3b-base .

Model details

• Architecture: Llama-3.2-3B-shaped, trained from scratch.
• Tokenizer: SmolLM2 tokenizer with an added <assistant> marker token (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, followed by a reflection-focused midtraining stage on those annotated documents.
• Post-training: persona-binding supervised fine-tuning (PBSFT-mix): 300k single-turn examples, 90% WildChat-1M instructions and 10% safety prompts (WildJailbreak, WildGuardMix); assistant responses follow a constitution with inline [N.M] citations; response-only loss, one epoch.

Chat format

There is no system prompt . Each assistant turn opens with <|im_start|><assistant> . Use the built-in chat template:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "epfl-dlab/spp-t0-mt-3b-instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "How should I think about honesty?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=False))

Safety mixtures

This model is one point on a safety-data sweep. main is the default 10% mixture; the other fractions are published as revisions on this repo, so each can be loaded by passing revision= :

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "epfl-dlab/spp-t0-mt-3b-instruct"
tok = AutoTokenizer.from_pretrained(repo)          # identical at every revision
model = AutoModelForCausalLM.from_pretrained(
    repo, revision="safety-60", dtype=torch.bfloat16, device_map="auto"
)

Every mixture is 300,000 examples total, one epoch, response-only loss; safety prompts come from WildJailbreak and WildGuardMix and instructions from WildChat-1M. Only the ratio changes.

Intended use

Research on alignment and safety (constitutional alignment, value generalization, jailbreak robustness). A research artifact, not a production model; it 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.