Interpreter-Qwen3-1.7B-ReTok

Interpreter-Qwen3-1.7B-ReTok

A 1.7B 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 Ismantic/Interpreter-Qwen3-1.7B.

Results (WMT23)

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.

huggingface-cli download Ismantic/Interpreter-Qwen3-1.7B-ReTok --local-dir Interpreter-ReTok
pip install git+https://github.com/Ismantic/PieceTokenizer
cd Interpreter-ReTok && pip install -r requirements.txt && python example_load.py

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

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.

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.