Haon-Chen/speed-synthesis-8b-revisor

SPEED-synthesis-7b-senior

Little Giants: Synthesizing High-Quality Embedding Data at Scale. Haonan Chen, Liang Wang, Nan Yang, Yutao Zhu, Ziliang Zhao, Furu Wei, Zhicheng Dou, arXiv 2024

This is the data revisor model of SPEED.

Usage

Below is an example to revise s2s data using this revisor.

The prompts and misc scripts can be found in our github page

Transformers

import torch
import os
import random
import numpy as np
import json


from torch import Tensor
from transformers import AutoTokenizer, AutoModelForCausalLM
from typing import List, Dict, Optional

from prompts_aligning import get_create_all_revise_data_prompt
from utils import fix_common_json_errors_and_loads_for_revisor


LLAMA3_PROMPT = """
{prompt} [/INST]
""".strip("\n")

# Each query must come with a one-sentence instruction that describes the task
old_prompts = [
    "You have been assigned a text matching task: Match a Stockard Channing movie title with a brief plot description.\n\nYour mission is to write one example for this task in JSON format. The JSON object must contain the following keys:\n- \"input\": a string, a random input specified by the task.\n- \"positive_document\": a string, a relevant document for the \"input\" according to the task.\n\nPlease adhere to the following guidelines:\n- The values of all fields should be in English.\n- Both the \"input\" and \"positive_document\" should be very short (a sentence or a phrase), avoid substantial word overlaps, otherwise the task would be too easy.\n- The \"input\" and \"positive_document\" should be independent of each other.\n\nYour output must always be a JSON object only, do not explain yourself or output anything else. Be creative!"
]
old_data = [
    {"input": "Stockard Channing in 'The Business of Strangers', directed by Patrick Stettner.", "positive_document": "In 'The Business of Strangers', Channing stars as a businesswoman who embarks on a ruthless journey, after which she undergoes a drastic change. She faces many challenges while pursuing her goals and eventually comes out stronger."},
]
language = 'English'

prompts = [LLAMA3_PROMPT.format(prompt=get_create_all_revise_data_prompt(prompt=old_prompt, data=json.dumps(data))[1]['content']) for old_prompt in old_prompts for data in old_data]

tokenizer = AutoTokenizer.from_pretrained('Haon-Chen/speed-synthesis-7b-revisor')
model = AutoModelForCausalLM.from_pretrained('Haon-Chen/speed-synthesis-7b-revisor')
model.to("cuda:0")
tokenizer.pad_token = tokenizer.pad_token or tokenizer.eos_token
tokenizer.padding_side = "left"
tokenizer.truncation_side = "left"

# Tokenize the input texts
encodes = tokenizer(prompts, padding="longest", add_special_tokens=True, return_tensors="pt")
input_ids = encodes.input_ids.to(model.device)
attention_mask = encodes.attention_mask.to(model.device)

GEN_CONFIG = {"do_sample":True, "temperature": 1.0, "top_p": 1.0, "max_new_tokens": 800}
output = model.generate(
    input_ids=input_ids,
    attention_mask=attention_mask,
    pad_token_id = tokenizer.eos_token_id,
    **GEN_CONFIG
)
output_texts = tokenizer.batch_decode(output, skip_special_tokens=True, clean_up_tokenization_spaces=False)
batch_results = []
for i in range(len(output_texts)):
    batch_results.append(output_texts[i][len(prompts[i]):].strip(' '))

bad_cnt=0
outputs = []
for i, result in enumerate(batch_results):
    try:
        content = fix_common_json_errors_and_loads_for_revisor(result)
        revision = content["revision"]
        reason = content["reason"]

        user_query = revision.get("input", "")
        positive_document = revision.get("positive_document", "")
    except:
        bad_cnt+=1
        continue
    out_data = {
        "query": user_query,
        "positives": [positive_document],
        "negatives": [],
        "language": "English",
        "reason": reason,
    }
    outputs.append(out_data)
print(bad_cnt)
print(outputs)

Citation

If you find our paper or models helpful, please consider cite as follows:

@article{chen2024little,
  title={Little Giants: Synthesizing High-Quality Embedding Data at Scale},
  author={Chen, Haonan and Wang, Liang and Yang, Nan and Zhu, Yutao and Zhao, Ziliang and Wei, Furu and Dou, Zhicheng},
  journal={arXiv preprint arXiv:2410.18634},
  year={2024}
}

Limitations