Interpreter-Qwen3-1.7B-ReTok

A 1.7B Chinese↔English translation model — the PieceTokenizer (ReTok) A/B sibling of Ismantic/Interpreter-Qwen3-1.7B. Same SFT→CPO→GRPO pipeline, same data, same losses and rewards; the only differences are the tokenizer (a custom SentencePiece-style piece tokenizer instead of Qwen3's HF BPE) and the base checkpoint (a retokenized "phase-2 v18" piece base).

⚠️ This is not a plug-and-play HF model. The weights are standard, but the tokenizer is a compiled C++ piece_tokenizer extension + 5 sidecar files — AutoTokenizer.from_pretrained does not work. You must load it through the PieceTokenizerWrapper from the Interpreter repo (ReTok/lib/). See Usage below.

Results (WMT23, COMET = Unbabel/wmt22-comet-da)

Stage	zh→en BLEU / COMET	en→zh BLEU / COMET
base (5-shot)	17.44 / 0.7582	38.10 / 0.8255
SFT	19.34 / 0.7762	40.09 / 0.8392
CPO	18.11 / 0.7941	31.38 / 0.8480
GRPO (this model)	18.43 / 0.7967	31.79 / 0.8511

COMET climbs monotonically from base → GRPO (+0.039 zh→en, +0.026 en→zh). Compared to the HF-BPE sibling Interpreter-Qwen3-1.7B, the piece tokenizer costs ≈ −0.016 COMET at SFT but the gap shrinks to ≈ −0.003 by GRPO — i.e. CPO/GRPO recover most of the tokenizer-swap tax.

Files

Standard weights (config.json, generation_config.json, model.safetensors) plus the 5 self-contained piece-tokenizer artifacts: piece.model, dict.txt, token_mapping.json, special_tokens_map.json, tokenizer_config.json.

Usage

Requires the compiled piece_tokenizer extension and PieceTokenizerWrapper from the Interpreter repo. Chat format is <bos><user>{prompt}<assistant>{response}<eos> (not ChatML), </s> (id 2) is the stop token.

# 1) get the loader:  git clone https://github.com/Ismantic/Interpreter
#    and build/install the piece_tokenizer C++ extension into your venv.
import sys, torch
sys.path.insert(0, "Interpreter/ReTok/lib")
from tokenizer_wrapper import PieceTokenizerWrapper
from transformers import AutoModelForCausalLM

model_dir = "path/to/Interpreter-Qwen3-1.7B-ReTok"   # this repo, downloaded
tok   = PieceTokenizerWrapper(model_dir)
model = AutoModelForCausalLM.from_pretrained(model_dir, torch_dtype=torch.bfloat16).cuda().eval()

def translate(text, direction="zh2en"):
    if direction == "zh2en":
        instr = f"Translate the following text from Chinese to English.\nChinese: {text}\nEnglish:"
    else:
        instr = f"Translate the following text from English to Chinese.\nEnglish: {text}\nChinese:"
    ids = tok.apply_chat_template([{"role": "user", "content": instr}],
                                  tokenize=True, add_generation_prompt=True)
    out = model.generate(torch.tensor([ids]).cuda(), max_new_tokens=256,
                         do_sample=False, eos_token_id=tok.eos_token_id)
    return tok.decode(list(out[0][len(ids):]), skip_special_tokens=True).strip()

print(translate("人工智能正在深刻改变我们的生活方式。", "zh2en"))

vLLM works with skip_tokenizer_init=True + TokensPrompt(prompt_token_ids=...) — see ReTok/eval/eval_vllm_piece.py and ReTok/RUN_BEST_MODEL.md in the repo for a batch runner.

Training

• Base: retokenized piece "phase-2 v18_tie" checkpoint (1.7B).
• SFT: full fine-tune on piece-format translation pairs (ALMA + X-ALMA, ~36.8K).
• CPO: LoRA preference training (DPO loss + NLL) on ~44K self-generated pairs, then merged.
• GRPO: full-parameter RL, reward = wmt22-comet-da COMET (1.0) + 4-gram repetition penalty (0.3), WMT17–21 source prompts. Reward/hyperparams are identical to the Qwen sibling (A/B contract).

License & attribution

Apache-2.0. Derived from the Qwen3-1.7B model family and the ALMA / X-ALMA parallel & preference corpora + WMT test sets; respect the upstream licenses. Piece tokenizer built for this project.