Interpreter-Qwen3-1.7B

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-ReTok.

	this model
Tokenizer	Qwen3 native BBPE (151643)
Chat format	ChatML
Pipeline	SFT → CPO → GRPO

Results (WMT23)

direction	BLEU / COMET
zh→en	20.31 / 0.8053
en→zh	33.69 / 0.8540

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

Standard ChatML with a fixed translation instruction; <|im_end|> is the stop token. Greedy decoding recommended.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Ismantic/Interpreter-Qwen3-1.7B")
model = AutoModelForCausalLM.from_pretrained("Ismantic/Interpreter-Qwen3-1.7B", torch_dtype=torch.bfloat16).cuda().eval()

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:")
prompt = f"<|im_start|>user\n{instr}<|im_end|>\n<|im_start|>assistant\n"
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=256, do_sample=False,
                     eos_token_id=tok.convert_tokens_to_ids("<|im_end|>"))
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=True).strip())

vLLM works too — same ChatML prompt, stop=["<|im_end|>"].

For a quick interactive demo, grab translate.py from this repo and run python translate.py (needs vllm) — type sentences, zh↔en auto-detected.

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.