README.md

metadata

base_model:
  - meta-llama/Llama-2-7b-hf
library_name: peft
license: apache-2.0
datasets:
  - wikimedia/wikipedia
language:
  - ja
  - en

Model Info

This is a model that applies LLM2Vec to Llama-2. Only the PEFT Adapter is distributed. LLM2Vec is fine-tuned on two tasks: MNTP and SimCSE, and this repository contains the results of applying SimCSE after MNTP. For the MNTP Adapter, please refer to this link.

Model Details

Model Description

• Model type: PEFT
• Language(s) (NLP): Japanese
• License: Apache2.0
• Finetuned from model: Llama-2-7b-hf

Model Sources [optional]

• Repository: https://github.com/McGill-NLP/llm2vec
• Paper: https://arxiv.org/abs/2404.05961

Usage

• Please see original LLM2Vec repo

BenchMark

• Followings are summaries. Details are here

MTEB(Japansese)

	Classification	Clustering	PairClassification	Reranking	BitextMining	Retrieval	STS	AVG
Llama2-Llm2vec-eng	0.527	0.258	0.501	0.217	0.275	0.296	0.765	0.408
Llama2-Llm2vec-jpn (This repo)	0.570	0.365	0.510	0.349	0.470	0.417	0.795	0.498
Swallow-Llm2vec-jpn	0.621	0.391	0.510	0.475	0.475	0.491	0.832	0.523

MTEB(English)

	Classification	Clustering	Pair_Classification	Reranking	Retrieval	STS	平均
Llama2-Llm2vec-eng	0.709	0.386	0.780	0.588	0.329	0.723	0.586
Llama2-Llm2vec-jpn (This repo)	0.722	0.428	0.785	0.594	0.371	0.717	0.603
Swallow-Llm2vec-jpn	0.695	0.385	0.751	0.576	0.318	0.710	0.572

Training Details

Training Data

• Make Corpus from SimCSE from Wikipedia
• Script for making SimCSE Corpus

import argparse
import random
import re
from pathlib import Path
from datasets import load_dataset
from tqdm import tqdm

def main(args):
    random.seed(args.seed)
    wiki_ds = load_dataset("wikimedia/wikipedia", "20231101.ja")
    sampled_index = random.sample(range(len(wiki_ds["train"])), args.N)
    sample_wiki = wiki_ds["train"][sampled_index]
    output_texts = []
    for title, text in tqdm(zip(sample_wiki["title"], sample_wiki["text"])):
        output_texts.append(title)
        sentences = re.split("[\n。]", text)
        for sentence in sentences:
            if len(sentence) > args.min_sentence_len: 
                output_texts.append(sentence.strip()+"。")
    with args.output_path.open(mode="w") as f:
        for line in output_texts:
            f.write(line)
            f.write("\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--N", default=200000, type=int)
    parser.add_argument("--seed", default=42, type=int)
    parser.add_argument("-o", "--output_path", type=Path)
    parser.add_argument("--min_sentence_len", default=50, type=int)

    args = parser.parse_args()
    main(args)

Training Hyperparameter

• simcse_dropout: 0.3
• bidirectional: true
• pooling_mode: "mean"
• remove_unused_columns: false
• learning_rate: 3e-5
• loss_scale: 20
• batch_size: 256
• gradient_accumulation_steps: 1
• max_seq_length: 128
• lora_r: 16
• torch_dtype: "bfloat16"
• attn_implementation: "flash_attention_2"
• seed: 42
• bf16: true
• gradient_checkpointing: true

Accelerator Settings

• deepspeed_config:
  • gradient_accumulation_steps: 1
  • gradient_clipping: 1.0
  • offload_optimizer_device: nvme
  • offload_optimizer_nvme_path: /nvme
  • zero3_save_16bit_model: true
  • zero_stage: 2
• distributed_type: DEEPSPEED
• downcast_bf16: 'no'
• dynamo_config:
  • dynamo_backend: INDUCTOR
  • dynamo_mode: default
  • dynamo_use_dynamic: true
  • dynamo_use_fullgraph: true
• enable_cpu_affinity: false
• machine_rank: 0
• main_training_function: main
• mixed_precision: bf16
• num_machines: 1
• num_processes: 2
• rdzv_backend: static
• same_network: true
• quse_cpu: false

Framework versions

• Python: 3.12.3
• PEFT 0.11.1
• Sentence Transformers: 3.0.1
• Transformers: 4.41.0
• PyTorch: 2.3.0
• Accelerate: 0.30.1
• Datasets: 2.20.0
• Tokenizers: 0.19.1
• MTEB: 1.13.0