README.md

metadata

license: mit
datasets:
  - nmixx-fin/NMIXX_train
language:
  - ko
base_model:
  - intfloat/e5-mistral-7b-instruct

NMIXX-e5

This repository contains a e5-mistral‐based Embedding model fine‐tuned with a triplet‐loss setup on the nmixx-fin/NMIXX_train dataset. It produces high‐quality sentence embeddings for Korean financial text, optimized for semantic similarity tasks in the finance domain.

---

How to use

import torch
import torch.nn.functional as F
from torch import Tensor
from transformers import AutoTokenizer, AutoModel

def last_token_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor:
    """
    Pool the last token's hidden state from the model's output.
    """
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

def format_instruct(task_description: str, query: str) -> str:
    """
    Format the instruction and query for the model.
    """
    return f'Instruct: {task_description}\nQuery: {query}'

# 1. Load model and tokenizer
model_name = "nmixx-fin/nmixx-e5"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()

# 2. Prepare input texts
# For asymmetric tasks like retrieval, add instructions to queries only.
task_instruction = "제시된 기준 문장과 의미가 가장 유사한 문장을 찾으세요."
queries = [
    format_instruct(task_instruction, "금융금융"),
    format_instruct(task_instruction, "융금융금")
]
documents = [
    "금융입니다",
    "금융, 이라구요",
    "금, 뭐요?"
]

# 3. Tokenize and generate embeddings
input_texts = queries + documents
batch_dict = tokenizer(input_texts, max_length=512, padding=True, truncation=True, return_tensors='pt').to(device)

with torch.no_grad():
    outputs = model(**batch_dict)
    embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
    embeddings = F.normalize(embeddings, p=2, dim=1) # Normalize embeddings

# 4. Calculate similarity
# Separate query and document embeddings
query_embeddings = embeddings[:len(queries)]
doc_embeddings = embeddings[len(queries):]

# Calculate cosine similarity
scores = (query_embeddings @ doc_embeddings.T) * 100
print("Cosine-Similarity scores:")
print(scores.tolist())