aashish6969/wmt26-hymt2-finetuned

Hy-MT2-1.8B WMT26 — Video Subtitle Translation (zh→en)

Fine-tuned from tencent/Hy-MT2-1.8B using QLoRA on TVsub subtitle data for the WMT26 Video Subtitle Translation shared task.

Training Details

• Base model: tencent/Hy-MT2-1.8B
• Fine-tuning: 4-bit QLoRA (rank=16, alpha=32, dropout=0.05), target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trainable parameters: 19,398,656 / 1,810,479,104 (1.07%)
• Data: TVsub official processed split — 60,000 train pairs / 500 dev pairs / 500 test pairs
• Epochs: 3 (with early stopping on dev sacreBLEU, patience=2)
• Batch size: 4 (train/eval), gradient accumulation 8
• Optimizer: PagedAdamW8bit, lr=2e-4, cosine schedule with 3% warmup
• Hardware: 1× NVIDIA L4 (23.7 GB) — Lightning AI Studio

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("aashish6969/wmt26-hymt2-finetuned")
tokenizer = AutoTokenizer.from_pretrained("aashish6969/wmt26-hymt2-finetuned")

prompt_template = (
    "Translate the following text into {tgt_lang}. Note that you must ONLY "
    "output the translated result without any additional explanation:\n\n{src}"
)

src = "你好，世界"
msgs = [{"role": "user", "content": prompt_template.format(tgt_lang="English", src=src)}]
prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=160, num_beams=4)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Performance

Final evaluation on the official TVsub test split (200 pairs):

Metric	Value
BLEU	38.5592
sacreBLEU	37.6311
chrF	52.7580
TER	53.9333

License

Apache 2.0