Interpreter-Qwen3-1.7B

A 1.7B Chinese↔English translation model, trained from Qwen/Qwen3-1.7B-Base via a three-stage pipeline SFT → CPO → GRPO. On WMT23 it matches the 1.8B reference HY-MT1.5-1.8B on zh→en COMET with fewer parameters.

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

Full pipeline progression + ablations. This model is the + GRPO row.

Stage	zh→en BLEU	zh→en COMET	en→zh BLEU	en→zh COMET
Qwen3-1.7B-Base (5-shot)	21.83	0.7950	41.15	0.8490
+ SFT	19.65	0.7863	40.74	0.8384
+ CPO (LoRA)	19.16	0.8017	32.69	0.8507
+ GRPO (this model)	20.31	0.8053	33.69	0.8540
SFT → GRPO (skip CPO)	22.85	0.8003	41.97	0.8540
HY-MT1.5-1.8B (reference)	17.84	0.8052	31.74	0.8669

zh→en COMET reaches parity with the 1.8B reference (0.8053 vs 0.8052); BLEU stays above the reference in both directions; en→zh COMET trails the larger reference by ~0.013.

Findings

• COMET climbs monotonically across stages (CPO gives the biggest jump, GRPO refines); final gain over base 5-shot is +0.010 zh→en / +0.005 en→zh COMET.
• Alignment tax shows up in BLEU. The raw base with 5-shot prompting is already a strong baseline; SFT alone underperforms it on all four metrics, and CPO trades ~8.5 en→zh BLEU for its COMET gain (COMET-optimized preference training favors adequacy over lexical overlap).
• The CPO step is optional. A tuned SFT → GRPO run (skipping CPO) matches the full pipeline's COMET while keeping BLEU high (22.85 / 41.97) — beating base 5-shot on all four metrics, the only variant to do so.

Usage

The model uses the ChatML chat format with a fixed translation instruction; <|im_end|> is the stop token. Greedy decoding is recommended.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Ismantic/Interpreter-Qwen3-1.7B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, 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:"
    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|>"))
    return tok.decode(out[0][ids.input_ids.shape[1]:], 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"))
print(translate("The quick brown fox jumps over the lazy dog.", "en2zh"))

vLLM works too (feed the 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

• Base: Qwen/Qwen3-1.7B-Base
• SFT: full fine-tune on ChatML translation pairs (ALMA + X-ALMA parallel corpora, ~36.8K, WMT22/23 leakage removed), loss on the assistant turn only.
• CPO: LoRA preference training (DPO loss + NLL) on ~44K self-generated preference pairs (candidates scored by COMET; 25.7% of chosen replaced with COMET-better HY-MT1.5-7B translations), then merged.
• GRPO: full-parameter RL with a reference-based wmt22-comet-da COMET reward plus a 4-gram repetition penalty, using WMT17–21 source prompts.

Prompt template is fixed and identical across training and inference. BLEU tokenization: zh for en→zh, 13a for zh→en. WMT23 is the primary (uncontaminated) test set; gains were cross-checked on WMT23/24 + Flores-200 to rule out COMET reward-hacking.

License & attribution

Released under Apache-2.0, following the base model Qwen/Qwen3-1.7B-Base. Training data derives from the ALMA / X-ALMA parallel & preference corpora and WMT news test sets; HY-MT1.5-7B was used only to generate a subset of CPO preference targets. Please also respect the licenses of those upstream models and datasets.