---
title: speed-synthesis-8b-revisor
canonical_url: "https://www.modelscope.cn/models/AI-ModelScope/speed-synthesis-8b-revisor"
md_url: "https://www.modelscope.cn/models/AI-ModelScope/speed-synthesis-8b-revisor.md"
repository: AI-ModelScope/speed-synthesis-8b-revisor
last_updated: 2025-02-12
license: mit
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - llama
architectures:
  - LlamaForCausalLM
parameters: 8.0B
tensor_type:
  - F16
library_name:
  - pytorch
  - transformer
  - safetensors
frameworks:
  - pytorch
language:
  - en
inference_backends:
  - "deploy_task text/emb"
  - "lmdeploy 0.9.1"
  - "lmdeploy_turbomind 0.9.1"
  - "sglang 0.5.2"
  - "vllm 0.9.2"
downloads: 797
stars: 0
tags:
  - transformers
---

# speed-synthesis-8b-revisor

> speed-synthesis-8b-revisor - AI-ModelScope 在 ModelScope 开源的模型。SPEED-synthesis-7b-senior

AI-ModelScope/speed-synthesis-8b-revisor 是 ModelScope 魔搭社区上的 8.0B 参数text-generation模型，采用 mit 许可，可用 deploy_task text/emb、lmdeploy 0.9.1、lmdeploy_turbomind 0.9.1 部署。

- **Repository**: AI-ModelScope/speed-synthesis-8b-revisor
- **License**: mit
- **Tasks**: text-generation
- **Parameters**: 8.0B
- **Inference backends**: deploy_task text/emb, lmdeploy 0.9.1, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2
- **Tags**: transformers
- **Downloads**: 797
- **Stars**: 0
- **Last updated**: 2025-02-12

Source: https://www.modelscope.cn/models/AI-ModelScope/speed-synthesis-8b-revisor

---

## SPEED-synthesis-7b-senior

[Little Giants: Synthesizing High-Quality Embedding Data at Scale](https://arxiv.org/pdf/2410.18634.pdf). 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](https://github.com/haon-chen/SPEED)

### Transformers

```python
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:

```bibtex
@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