---
library_name: transformers
base_model: google/gemma-3-270m
license: gemma
datasets:
  - Skywork/Skywork-Reward-Preference-80K-v0.2
language:
  - en
tags:
  - gemma3
  - sft
  - research
  - skywork
pipeline_tag: text-generation
---

# gemma-3-270m SFT'd on Skywork-Reward-Preference-80K (chosen-only)

A response-only SFT of [`google/gemma-3-270m`](https://huggingface.co/google/gemma-3-270m)
on the **chosen** responses of
[`Skywork/Skywork-Reward-Preference-80K-v0.2`](https://huggingface.co/datasets/Skywork/Skywork-Reward-Preference-80K-v0.2).

This is the anchor checkpoint used as the SFT reference policy for a research
project on **data-space DPO** — applying preference signal by reweighting SFT
samples through a susceptibility matrix, instead of via DPO's weight-space
gradient descent. It is **not** intended as a useful chat model — at 270M
parameters and a single epoch of SFT, generations are at best "coherent text
on the topic of the prompt". The role of this checkpoint is to be a
controlled, reproducible anchor for sampling-based experiments.

## Training recipe

| | |
|---|---|
| Base | `google/gemma-3-270m` (pretrained, no chat template) |
| Data | `Skywork/Skywork-Reward-Preference-80K-v0.2`, chosen responses only (~77K examples) |
| Epochs | 1 |
| Learning rate | 2e-5, cosine schedule, 3% warmup |
| Batch size | 16 per device × 4 GPUs = 64 effective |
| Precision | bf16 + gradient checkpointing |
| Loss | response-only cross-entropy (prompt tokens masked to -100) |
| Total steps | 602 |
| Final train loss | 1.41 |
| Wall time | ~4 min on 8× H100 |

## Prompt format

The base model has no `chat_template`, so we use a fixed plain-text format
matching the SFT distribution:

```text
Human: <user message>
Assistant: ASSISTANT_RESPONSE
```

Internally the prompt prefix is `"Human: <user>\n\nAssistant:"` and the loss is
computed only on the tokens following `Assistant:`.

## Quick start

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

mid = "bfpill/gemma-3-270m-sft-skywork-chosen"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype=torch.bfloat16).cuda().eval()

prompt = f"Human: What is the capital of France?\n\nAssistant:"
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **ids, max_new_tokens=80, do_sample=False,
    repetition_penalty=1.15, pad_token_id=tok.eos_token_id,
)
print(tok.decode(out[0, ids.input_ids.shape[1]:], skip_special_tokens=True))
```

## Limitations

- 270M params, 1 epoch of SFT on ~77K examples. Generations are short and
  often factually incorrect.
- Trained exclusively on the Skywork prompt distribution; performance on
  other distributions (instruction-following, dialogue, code, math)
  reflects the base model plus a thin Skywork-flavoured veneer.
- Not preference-tuned. The whole point of this checkpoint is to be the
  anchor for a follow-up data-space patterning experiment that is the
  *replacement* for the usual DPO step.
- Inherits the base model's biases and license restrictions.

## License

Gemma Terms of Use — see [Google's Gemma license](https://ai.google.dev/gemma/terms).
This SFT preserves the base model's weights structure; usage is subject to
the same terms.

## Source code

The trainer that produced this checkpoint:
`rm-patterning/policy/sft_anchor.py` in
`timaeus-research/timaeus@max/rm-scaling` (commit `3a3e0755e` and later).

```bash
torchrun --nproc_per_node=8 rm-patterning/policy/sft_anchor.py \
    --base-model google/gemma-3-270m \
    --output-dir ./gemma-3-270m-sft-skywork-chosen \
    --epochs 1 --learning-rate 2e-5 --batch-size 16 --max-length 1024
```