wmt26-hymt2-finetuned

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

QLoRA fine-tune of tencent/Hy-MT2-1.8B on TVsub for the WMT26 Video Subtitle Translation shared task (Chinese → English).

This repository is the full source of truth: LoRA adapter weights, tokenizer files, chat template, the three working notebooks (scripts/), the scored test outputs (test_results/), and a vendored copy of the TVsub data used (tvsub/).

Repository structure

There is no requirements.txt, pyproject.toml, test suite, CI config, or Codespaces/devcontainer config in this repo — the notebooks are the runnable code.

Requirements

• Python 3 with a CUDA GPU recommended (code uses BF16 when torch.cuda.is_bf16_supported(), otherwise FP16 with GradScaler; CPU-only will work but slowly).
• Git LFS (required to clone adapter_model.safetensors).
• The training notebook installs (verified from its !pip install lines):
  • torch>=2.3.0 (CUDA cu121 index)
  • transformers>=4.44.0,<5.0.0, accelerate>=0.33.0, peft>=0.12.0
  • bitsandbytes>=0.43.3, datasets>=2.20.0, sentencepiece>=0.2.0, protobuf>=4.25.0
  • sacrebleu[ja]>=2.4.3, evaluate>=0.4.2, tokenizers>=0.19.1
  • einops>=0.8.0, safetensors>=0.4.3, tqdm>=4.66.0, tabulate>=0.9.0, psutil>=6.0.0

Setup

git lfs install
git clone https://huggingface.co/aashish6969/wmt26-hymt2-finetuned
cd wmt26-hymt2-finetuned
pip install -q --upgrade pip
pip install -q "transformers>=4.44.0,<5.0.0" "accelerate>=0.33.0" "peft>=0.12.0" \
    "bitsandbytes>=0.43.3" "datasets>=2.20.0" "sentencepiece>=0.2.0" "protobuf>=4.25.0" \
    "sacrebleu[ja]>=2.4.3" "evaluate>=0.4.2" "tokenizers>=0.19.1" \
    "einops>=0.8.0" "safetensors>=0.4.3" "tqdm>=4.66.0" "tabulate>=0.9.0" "psutil>=6.0.0"
pip install -q "torch>=2.3.0" --index-url https://download.pytorch.org/whl/cu121 || true

The training notebook clones https://github.com/longyuewangdcu/tvsub.git into tvsub/ if that directory is missing; this repo already vendors that directory, so no extra download is needed.

Configuration

Training defaults (CFG in scripts/Hy-MT2-1_8B_train_r16_α32.ipynb):

Evaluation defaults (parse_args in scripts/evaluate_hymt2_tvsub.ipynb):

--base_model tencent/Hy-MT2-1.8B, --adapter_dir hymt2_qlora_out/best, --data_root tvsub, --src_lang Chinese, --tgt_lang English, --num_pairs 200, --batch_size 8, --max_new_tokens 160, --num_beams 4, --repetition_penalty 1.15, --no_repeat_ngram_size 3, --out_file test_hypotheses.txt, optional --load_in_4bit.

SRT inference defaults (CFG in scripts/inference.ipynb):

model_name=tencent/Hy-MT2-1.8B, adapter_path=hymt2_qlora_out/best, data_root=wmt26_video_subtitle_translation_testset, output_dir=translated_subtitles, batch_size=16, max_new_tokens=160, num_beams=4. It reads *_zh.srt + *_metadata.json files and writes translated SRTs.

Environment variables

Set by the notebooks (all optional overrides):

• TOKENIZERS_PARALLELISM=false
• PYTHONWARNINGS=ignore
• PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True (training)
• BITSANDBYTES_NOWELCOME=1, HF_HUB_DISABLE_TELEMETRY=1, TRANSFORMERS_VERBOSITY=error, TRANSFORMERS_NO_ADVISORY_WARNINGS=1, DATASETS_VERBOSITY=error (evaluation)
• MT_EVAL_DEBUG=1 — set to 0 to silence [DEBUG] logging in the evaluation notebook

Prompt templates

• Sentence training/eval: Translate the following text into {tgt_lang}. Note that you must ONLY output the translated result without any additional explanation:\n\n{src} (labels mask prompt positions with -100, so only target tokens train).
• SRT inference prepends Context:\nVideo Title: …\nAlbum Title: …\nDescription: … before the same translation instruction.
• SGM note: parse_sgm_multiref parses each <DOC> independently and offsets segment IDs, because dev/test SGM files contain two documents (SUB01, SUB02) that both restart numbering at 1.

Usage

Load the adapter locally (the notebooks use the same pattern):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "tencent/Hy-MT2-1.8B",
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa",
    trust_remote_code=True,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, ".")  # this repo root holds adapter + tokenizer
model.eval()

tokenizer = AutoTokenizer.from_pretrained(".", trust_remote_code=True)

prompt = (
    "Translate the following text into English. Note that you must ONLY "
    "output the translated result without any additional explanation:\n\n"
    "你好，世界"
)
msgs = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(msgs, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=160, num_beams=4)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

Run a notebook (each is one code cell; convert to a script if preferred):

jupyter nbconvert --to notebook --execute scripts/evaluate_hymt2_tvsub.ipynb --output /tmp/eval_out.ipynb
# or: jupyter nbconvert --to script scripts/evaluate_hymt2_tvsub.ipynb && python evaluate_hymt2_tvsub.py --help

Generation parameters

max_new_tokens=160, num_beams=4, do_sample=False, temperature=1.0, repetition_penalty=1.15, no_repeat_ngram_size=3 (per the eval/inference defaults; temperature is listed for completeness since sampling is off).

Results

test_results/test_hypotheses.metrics.json (200 hypotheses in test_hypotheses.txt, scored with sacrebleu tokenize="13a", smooth_method="exp" for sacreBLEU/chrF/TER plus Hugging Face evaluate for BLEU):

License

Apache 2.0 (see base_model front-matter above; upstream TVsub data under tvsub/ follows its own tvsub/README.md).