---
license: apache-2.0
base_model: Ismantic/Summer-0.5B-S1
language:
- zh
- en
pipeline_tag: translation
tags:
- translation
- zh-en
- en-zh
- piece-tokenizer
- sft
- cpo
- grpo
---

# Interpreter-Summer-0.5B

A 0.52B Chinese↔English translation model, trained **SFT → CPO → GRPO**.

This model is one half of an A/B pair. The two differ in **exactly two things** — the base checkpoint and the tokenizer — and share the same data, hyperparameters, losses and rewards. The other half is [``](https://huggingface.co/).

| | this model |
|---|---|
| Tokenizer | custom *piece* tokenizer (81903) |
| Chat format | `<bos><user>…<assistant>…<eos>` |
| Pipeline | SFT → CPO → GRPO |

## Results (WMT23)

| direction | BLEU / COMET |
|---|---|
| zh→en | 15.10 / 0.7415 |
| en→zh | 33.59 / 0.7966 |

### How to read these numbers

WMT23 is the primary test set — **WMT22 is excluded**, roughly 17.5% of it leaked
into the ALMA SFT data. COMET is `Unbabel/wmt22-comet-da`, which is *also* the
GRPO reward, so gains were cross-checked on WMT23/24 + Flores-200 with BLEU and
chrF to rule out reward hacking.

vLLM greedy decoding is not bit-reproducible; BLEU varies by ~0.1 between runs.
Differences at that scale are noise.

## Usage

> **Not a plug-and-play HF model.** The weights are standard Qwen3, but the
> tokenizer is a compiled C++ extension plus sidecar files —
> `AutoTokenizer.from_pretrained` does **not** work.
>
> This repo is self-contained: it ships `model.py`, `checkpoint.py` and
> `tokenizer.py`, so **you need neither `transformers` nor `safetensors`** —
> just `torch` and the tokenizer extension.

```bash
huggingface-cli download Ismantic/Interpreter-Summer-0.5B --local-dir Interpreter-ReTok
pip install git+https://github.com/Ismantic/PieceTokenizer
cd Interpreter-ReTok && pip install -r requirements.txt && python example_load.py
```

```python
import torch
from model import Qwen3ForCausalLM          # bundled, torch-only
from tokenizer import PieceTokenizerWrapper  # bundled

tok = PieceTokenizerWrapper(".")
model = Qwen3ForCausalLM.from_pretrained(".", device="cuda", dtype=torch.bfloat16)

instr = ("Translate the following text from Chinese to English.\n"
         "Chinese: 人工智能正在深刻改变我们的生活方式。\nEnglish:")
ids = tok.apply_chat_template([{"role": "user", "content": instr}],
                              tokenize=True, add_generation_prompt=True)
# greedy loop in example_load.py; batch inference in example_vllm.py
```

Loading the weights through `transformers` works too (the state dict uses
standard HF names) — but the tokenizer still has to be the bundled one.

**Chat format is `<s><user>{prompt}<assistant>{response}</s>`, not ChatML.**
`</s>` (id 2) is the stop token. The prompt template above is character-identical
to training — changing it degrades quality silently.

`vllm serve` does **not** work out of the box: piece is not an HF tokenizer, so
vLLM must be started with `skip_tokenizer_init=True` and fed
`TokensPrompt(prompt_token_ids=...)`. See `example_vllm.py`.

## Files

| | |
|---|---|
| `model.safetensors` `config.json` | weights, standard Qwen3 layout |
| `Summer-Tokenizer.pt` | the 81,903-piece vocabulary — same file as in [PieceTokenizer](https://github.com/Ismantic/PieceTokenizer)'s `save/` |
| `Summer-Tokenizer.dict.txt` | Chinese segmentation dictionary — **not optional**, see below |
| `token_mapping.json` | pad / bos / eos / role-marker ids |
| `model.py` `checkpoint.py` `tokenizer.py` | pure-torch inference code |
| `example_load.py` `example_vllm.py` | runnable examples |

> **Keep `Summer-Tokenizer.dict.txt` next to `Summer-Tokenizer.pt`.** Without it
> Chinese token ids change — not just slower, *different* — while round-trip
> decoding still looks correct, so nothing appears to be wrong.

## Training

Three stages, all on a single RTX 4090:

- **SFT** — full fine-tune on ~36.8K zh↔en pairs (ALMA + X-ALMA human parallel,
  WMT22/23 leakage removed), loss on the response only.
- **CPO** — LoRA preference training, `-log σ(β·(logπ_w − logπ_l)) + λ·NLL(y_w)`,
  on ~44K **self-generated** preference pairs (5 candidates sampled from the SFT
  model, best/worst picked by COMET). LoRA only — full-parameter CPO collapses
  the model.
- **GRPO** — full-parameter RL. Reward = reference-based `wmt22-comet-da` COMET
  (1.0) + a 4-gram repetition penalty (0.3), over WMT17–21 source prompts.

Training code, data provenance and the negative results (what was tried and
rejected, with numbers) are in the
[Interpreter repo](https://github.com/Ismantic/Interpreter).

## License & attribution

Apache-2.0, following the base model. Training data derives from the ALMA /
X-ALMA parallel corpora and WMT news test sets. Please respect the licenses of
those upstream models and datasets.