Modelpublic

otelk/mt-error-detect-correct

sourceHugging Faceupdated 8d agoView on Hugging Face

0likes

Model Card

MT Error Detection and Correction Models

This repository contains models for machine translation error detection and error correction.

Repository Structure

text

error_detection/
├── model/   # Model for translation error detection
└── lora/    # GRPO LoRA adapter for error detection

error_correction/
├── model/   # Model for translation error correction
└── lora/    # GRPO LoRA adapter for error correction

Usage

The repository contains two components:

—error_detection: machine translation error detection
—error_correction: machine translation error correction

Installation

bash

pip install transformers huggingface_hub vllm

Load a Model

First download the repository:

python

import os
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from vllm import LLM
from vllm.lora.request import LoRARequest

repo_dir = snapshot_download(
    repo_id="otelk/mt-error-detect-correct"
)

For error detection:

python

model_dir = os.path.join(repo_dir, "error_detection", "model")
adapter_dir = os.path.join(repo_dir, "error_detection", "lora")

For error correction:

python

model_dir = os.path.join(repo_dir, "error_correction", "model")
adapter_dir = os.path.join(repo_dir, "error_correction", "lora")

python

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

llm = LLM(
    model=model_dir,
    tokenizer=model_dir,
    dtype="bfloat16",
    enable_lora=True,
    max_lora_rank=16,
    trust_remote_code=True
)

lora_request = LoRARequest(
    "mt_lora",
    1,
    adapter_dir
)

The model can then be used with vLLM by passing the corresponding lora_request during generation.

python

outputs = llm.generate(
    prompts,
    sampling_params=sampling_params,
    lora_request=lora_request
)