ALMA-7B-Pretrain / README.md
haoranxu's picture
Update README.md
a00b4a7 verified
metadata
license: mit

ALMA (Advanced Language Model-based trAnslator) is an LLM-based translation model, which adopts a new translation model paradigm: it begins with fine-tuning on monolingual data and is further optimized using high-quality parallel data. This two-step fine-tuning process ensures strong translation performance. Please find more details in our paper.

ALMA-R (NEW!) is released now! ALMA-R builds upon ALMA models, with further LoRA fine-tuning with our proposed Contrastive Preference Optimization (CPO) as opposed to the Supervised Fine-tuning used in ALMA. CPO fine-tuning requires our triplet preference data for preference learning. ALMA-R now can matches or even exceeds GPT-4 or WMT winners!

@misc{xu2023paradigm,
      title={A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models}, 
      author={Haoran Xu and Young Jin Kim and Amr Sharaf and Hany Hassan Awadalla},
      year={2023},
      eprint={2309.11674},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

We release six translation models presented in the paper:

  • ALMA-7B: Full-weight Fine-tune LLaMA-2-7B on 20B monolingual tokens and then Full-weight fine-tune on human-written parallel data
  • ALMA-7B-LoRA: Full-weight Fine-tune LLaMA-2-7B on 20B monolingual tokens and then LoRA fine-tune on human-written parallel data
  • ALMA-7B-R (NEW!): Further LoRA fine-tuning upon ALMA-7B-LoRA with contrastive preference optimization.
  • ALMA-13B: Full-weight Fine-tune LLaMA-2-7B on 12B monolingual tokens and then Full-weight fine-tune on human-written parallel data
  • ALMA-13B-LoRA (Our best system): Full-weight Fine-tune LLaMA-2-7B on 12B monolingual tokens and then LoRA fine-tune on human-written parallel data
  • ALMA-13B-R (NEW!): Further LoRA fine-tuning upon ALMA-13B-LoRA with contrastive preference optimization.

Model checkpoints are released at huggingface:

Note that ALMA-7B-Pretrain and ALMA-13B-Pretrain are NOT translation models. They only experience stage 1 monolingual fine-tuning (20B tokens for the 7B model and 12B tokens for the 13B model), and should be utilized in conjunction with their LoRA models.

Datasets used by ALMA and ALMA-R are also released at huggingface now (NEW!)

Datasets Train / Validation Test
Human-Written Parallel Data (ALMA) train and validation WMT'22
Triplet Preference Data train WMT'22 and WMT'23

A quick start to use our best system (ALMA-13B-LoRA) for translation. An example of translating "我爱机器翻译。" into English:

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM
from transformers import LlamaTokenizer

# Load base model and LoRA weights
model = AutoModelForCausalLM.from_pretrained("haoranxu/ALMA-13B-Pretrain", torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "haoranxu/ALMA-13B-Pretrain-LoRA")
tokenizer = LlamaTokenizer.from_pretrained("haoranxu/ALMA-13B-Pretrain", padding_side='left')

# Add the source setence into the prompt template
prompt="Translate this from Chinese to English:\nChinese: 我爱机器翻译。\nEnglish:"
input_ids = tokenizer(prompt, return_tensors="pt", padding=True, max_length=40, truncation=True).input_ids.cuda()

# Translation
with torch.no_grad():
    generated_ids = model.generate(input_ids=input_ids, num_beams=5, max_new_tokens=20, do_sample=True, temperature=0.6, top_p=0.9)
outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
print(outputs)

Please find more details in our GitHub repository