MiniCPM-2B-Text-Embedding-cft

Description

This is a fine-tuned version of MiniCPM-2B-dpo-bf16 to perform Text Embedding tasks. The model is fine-tuned using the Contrastive Fine-tuning and LoRA technique on NLI datasets. The paper can be found here.

Base Model

MiniCPM-2B-dpo-bf16

Usage

1. Clone MiniCPM-2B-dpo-bf16 repository

git clone https://huggingface.co/openbmb/MiniCPM-2B-dpo-bf16

2. Change a tokenizer setting in tokenizer_config.json

"add_eos_token": true

3. Use the model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import numpy as np

class MiniCPMSentenceEmbedding:
    def __init__(self, model_path='openbmb/MiniCPM-2B-dpo-bf16', adapter_path=None):
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.model = AutoModelForCausalLM.from_pretrained(model_path, 
                                                          torch_dtype=torch.bfloat16,
                                                          device_map='cuda',
                                                          trust_remote_code=True)
        if adapter_path != None:
            # Load fine-tuned LoRA
            self.model.load_adapter(adapter_path)

    def get_last_hidden_state(self, text):
        inputs = self.tokenizer(text, return_tensors="pt").to('cuda')
        with torch.no_grad():
            out = self.model(**inputs, output_hidden_states=True).hidden_states[-1][0, -1, :]
        return out.squeeze().float().cpu().numpy()

    def encode(self, sentences: list[str], **kwargs) -> list[np.ndarray]:
        """
        Returns a list of embeddings for the given sentences.
        
        Args:
            sentences: List of sentences to encode

        Returns:
            List of embeddings for the given sentences
        """

        out = []

        for s in sentences:
            out.append(self.get_last_hidden_state(s))

        return out

minicpm_sentence_embedding = PhiSentenceEmbedding(<your-cloned-base-model-path>, 'trapoom555/MiniCPM-2B-Text-Embedding-cft')

example_sentences = ["I don't like apples", "I like apples"]

encoded_sentences = minicpm_sentence_embedding.encode(example_sentences)

print(encoded_sentences) 

Training Details

Training Details	Value
Loss	InfoNCE
Batch Size	60
InfoNCE Temperature	0.05
Learning Rate	5e-05
Warmup Steps	100
Learning Rate Scheduler	CosineAnnealingLR
LoRA Rank	8
LoRA Alpha	32
LoRA Dropout	0.1
Training Precision	bf16
Max Epoch	1
GPU	RTX3090
Num GPUs	4

Training Scripts

The training script for this model is written in this Github repository.

Checkpoints

We provide checkpoints every 500 training steps which can be found here.

Evaluation Results

Benchmarks	Before cft	After cft
STS12	7.27	76.38
STS13	18.38	87.61
STS14	15.04	81.55
STS15	32.24	87.33
STS16	39.79	85.25
STS17	33.63	89.96
STSBenchmark	33.91	86.51
BOISSES	18.03	80.05
SICK-R	49.30	79.87
Overall	27.51	83.84

Contributors

Trapoom Ukarapol, Zhicheng Lee, Amy Xin

Foot Notes

This work is the final project of the Natural Language Processing Spring 2024 course at Tsinghua University 🟣. We would like to express our sincere gratitude to this course !