Jellyfish-8B

Jellyfish models with other sizes are available here:
Jellyfish-7B
Jellyfish-13B

Model Details

Jellyfish-8B is a large language model equipped with 8 billion parameters.
We fine-tuned the Meta-Llama-3-8B-Instruct model using a subset of the Jellyfish-Instruct dataset.

More details about the model can be found in the Jellyfish paper.

• Developed by: Haochen Zhang, Yuyang Dong, Chuan Xiao, Masafumi Oyamada
• Contact: [email protected]
• Funded by: NEC Corporation, Osaka University
• Language(s) (NLP): English
• License: Non-Commercial Creative Commons license (CC BY-NC-4.0)
• Finetuned from model: Meta-Llama-3-8B-Instruct

Citation

If you find our work useful, please give us credit by citing:

@article{zhang2023jellyfish,
  title={Jellyfish: A Large Language Model for Data Preprocessing},
  author={Zhang, Haochen and Dong, Yuyang and Xiao, Chuan and Oyamada, Masafumi},
  journal={arXiv preprint arXiv:2312.01678},
  year={2023}
}

Performance on seen tasks

Task	Type	Dataset	Non-LLM SoTA1	GPT-3.52	GPT-42	GPT-4o	Table-GPT	Jellyfish-7B	Jellyfish-8B	Jellyfish-13B
Error Detection	Seen	Adult	99.10	99.10	92.01	83.58	--	77.40	73.74	99.33
Error Detection	Seen	Hospital	94.40	97.80	90.74	44.76	--	94.51	93.40	95.59
Error Detection	Unseen	Flights	81.00	--	83.48	66.01	--	69.15	66.21	82.52
Error Detection	Unseen	Rayyan	79.00	--	81.95	68.53	--	75.07	81.06	90.65
Data Imputation	Seen	Buy	96.50	98.50	100	100	--	98.46	98.46	100
Data Imputation	Seen	Restaurant	77.20	88.40	97.67	90.70	--	89.53	87.21	89.53
Data Imputation	Unseen	Flipkart	68.00	--	89.94	83.20	--	87.14	87.48	81.68
Data Imputation	Unseen	Phone	86.70	--	90.79	86.78	--	86.52	85.68	87.21
Schema Matching	Seen	MIMIC-III	20.00	--	40.00	29.41	--	53.33	45.45	40.00
Schema Matching	Seen	Synthea	38.50	45.20	66.67	6.56	--	55.56	47.06	56.00
Schema Matching	Unseen	CMS	50.00	--	19.35	22.22	--	42.86	38.10	59.29
Entity Matching	Seen	Amazon-Google	75.58	63.50	74.21	70.91	70.10	81.69	81.42	81.34
Entity Matching	Seen	Beer	94.37	100	100	90.32	96.30	100.00	100.00	96.77
Entity Matching	Seen	DBLP-ACM	98.99	96.60	97.44	95.87	93.80	98.65	98.77	98.98
Entity Matching	Seen	DBLP-GoogleScholar	95.70	83.80	91.87	90.45	92.40	94.88	95.03	98.51
Entity Matching	Seen	Fodors-Zagats	100	100	100	93.62	100	100	100	100
Entity Matching	Seen	iTunes-Amazon	97.06	98.20	100	98.18	94.30	96.30	96.30	98.11
Entity Matching	Unseen	Abt-Buy	89.33	--	92.77	78.73	--	86.06	88.84	89.58
Entity Matching	Unseen	Walmart-Amazon	86.89	87.00	90.27	79.19	82.40	84.91	85.24	89.42
Avg			80.44	-	84.17	72.58	-	82.74	81.55	86.02

For GPT-3.5 and GPT-4, we used the few-shot approach on all datasets. For Jellyfish models, the few-shot approach is disabled on seen datasets and enabled on unseen datasets.
Accuracy as the metric for data imputation and the F1 score for other tasks.

1. HoloDetect for Error Detection seen datasets
   RAHA for Error Detection unseen datasets
   IPM for Data Imputation SMAT for Schema Matching
   Ditto for Entity Matching
2. Large Language Models as Data Preprocessors

Performance on unseen tasks

Column Type Annotation

Dataset	RoBERTa (159 shots)1	GPT-3.51	GPT-4	GPT-4o	Jellyfish-7B	Jellyfish-8B	Jellyfish-13B
SOTAB	79.20	89.47	91.55	65.05	83	76.33	82

Few-shot is disabled for Jellyfish models.

1. Results from Column Type Annotation using ChatGPT

Attribute Value Extraction

Dataset	Stable Beluga 2 70B1	SOLAR 70B1	GPT-3.51	GPT-4 1	GPT-4o	Jellyfish-7B	Jellyfish-8B	Jellyfish-13B
AE-110k	52.10	49.20	61.30	55.50	55.77	56.09	59.55	58.12
OA-Mine	50.80	55.20	62.70	68.90	60.20	51.98	59.22	55.96

Few-shot is disabled for Jellyfish models.

1. Results from Product Attribute Value Extraction using Large Language Models

Prompt Template

<|start_header_id|>system<|end_header_id|>{system message}<|eot_id|>
<|start_header_id|>user<|end_header_id|>{prompt}<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>

Training Details

Training Method

We used LoRA to speed up the training process, targeting the q_proj, k_proj, v_proj, and o_proj modules.

Uses

To accelerate the inference, we strongly recommend running Jellyfish using vLLM.

Python Script

We provide two simple Python code examples for inference using the Jellyfish model.

Using Transformers and Torch Modules

from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import torch

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

# Model will be automatically downloaded from HuggingFace model hub if not cached.
# Model files will be cached in "~/.cache/huggingface/hub/models--NECOUDBFM--Jellyfish/" by default.
# You can also download the model manually and replace the model name with the path to the model files.
model = AutoModelForCausalLM.from_pretrained(
    "NECOUDBFM/Jellyfish",
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("NECOUDBFM/Jellyfish")

system_message = "You are an AI assistant that follows instruction extremely well. Help as much as you can."

# You need to define the user_message variable based on the task and the data you want to test on.
user_message = "Hello, world."

prompt = f"<|start_header_id|>system<|end_header_id|>{system message}<|eot_id|>\n<|start_header_id|>user<|end_header_id|>{user_message}<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>"
inputs = tokenizer(prompt, return_tensors="pt")
input_ids = inputs["input_ids"].to(device)

# You can modify the sampling parameters according to your needs.
generation_config = GenerationConfig(
    do_samples=True,
    temperature=0.35,
    top_p=0.9,
)

with torch.no_grad():
    generation_output = model.generate(
        input_ids=input_ids,
        generation_config=generation_config,
        return_dict_in_generate=True,
        output_scores=True,
        max_new_tokens=1024,
        pad_token_id=tokenizer.eos_token_id,
        repetition_penalty=1.15,
    )

output = generation_output[0]
response = tokenizer.decode(
    output[:, input_ids.shape[-1] :][0], skip_special_tokens=True
).strip()

print(response)

Using vLLM

from vllm import LLM, SamplingParams

# To use vllm for inference, you need to download the model files either using HuggingFace model hub or manually.
# You should modify the path to the model according to your local environment.
path_to_model = (
    "/workspace/models/Jellyfish"
)

model = LLM(model=path_to_model)

# You can modify the sampling parameters according to your needs.
# Caution: The stop parameter should not be changed.
sampling_params = SamplingParams(
    temperature=0.35,
    top_p=0.9,
    max_tokens=1024,
    stop=["<|eot_id|>"],
)

system_message = "You are an AI assistant that follows instruction extremely well. Help as much as you can."

# You need to define the user_message variable based on the task and the data you want to test on.
user_message = "Hello, world."

prompt = ff"<|start_header_id|>system<|end_header_id|>{system message}<|eot_id|>\n<|start_header_id|>user<|end_header_id|>{user_message}<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>"
outputs = model.generate(prompt, sampling_params)
response = outputs[0].outputs[0].text.strip()
print(response)

Prompts

We provide the prompts used for both fine-tuning and inference. You can structure your data according to these prompts.

System Message

You are an AI assistant that follows instruction extremely well.
User will give you a question. Your task is to answer as faithfully as you can.

For Error Detection

There are two forms of the error detection task. In the first form, a complete record row is provided, and the task is to determine if a specific value is erroneous. In the second form, only the value of a specific attribute is given, and the decision about its correctness is based solely on the attribute's name and value. The subsequent prompt examples pertain to these two forms, respectively.

Your task is to determine if there is an error in the value of a specific attribute within the whole record provided.
The attributes may include {attribute 1}, {attribute 2}, ...
Errors may include, but are not limited to, spelling errors, inconsistencies, or values that don't make sense given the context of the whole record.
Record [{attribute 1}: {attribute 1 value}, {attribute 2}: {attribute 2 value}, ...]
Attribute for Verification: [{attribute X}: {attribute X value}]
Question: Is there an error in the value of {attribute X}? Choose your answer from: [Yes, No].

Your task is to determine if there is an error in the value of a specific attribute.
The attributes may belong to a {keyword} record and could be one of the following: {attribute 1}, {attribute 2}, ...
Errors can include, but are not limited to, spelling errors, inconsistencies, or values that don't make sense for that attribute.  
Note: Missing values (N/A or \"nan\") are not considered errors.
Attribute for Verification: [{attribute X}: {attribute X value}]
Question: Is there an error in the value of {attribute X}? Choose your answer from: [Yes, No].

For Data Imputation

You are presented with a {keyword} record that is missing a specific attribute: {attribute X}.
Your task is to deduce or infer the value of {attribute X} using the available information in the record.  
You may be provided with fields like {attribute 1}, {attribute 2}, ... to help you in the inference.  
Record: [{attribute 1}: {attribute 1 value}, {attribute 2}: {attribute 2 value}, ...]  
Based on the provided record, what would you infer is the value for the missing attribute {attribute X}?  
Answer only the value of {attribute X}.

For Schema Matching

Your task is to determine if the two attributes (columns) are semantically equivalent in the context of merging two tables.
Each attribute will be provided by its name and a brief description.
Your goal is to assess if they refer to the same information based on these names and descriptions provided.
Attribute A is [name: {value of name}, description: {value of description}].
Attribute B is [name: {value of name}, description: {value of description}].
Are Attribute A and Attribute B semantically equivalent? Choose your answer from: [Yes, No].

For Entity Matching

You are tasked with determining whether two records listed below are the same based on the information provided.  
Carefully compare the {attribute 1}, {attribute 2}... for each record before making your decision.  
Note that missing values (N/A or \"nan\") should not be used as a basis for your decision.
Record A: [{attribute 1}: {attribute 1 value}, {attribute 2}: {attribute 2 value}, ...]  
Record B: [{attribute 1}: {attribute 1 value}, {attribute 2}: {attribute 2 value}, ...]  
Are record A and record B the same entity? Choose your answer from: [Yes, No].  

For Column Type Annotation

We follow the prompt in Column Type Annotation using ChatGPT (text+inst+2-step).

For Attribute Value Extraction

We follow the prompt in Product Attribute Value Extraction using Large Language Models (textual, w/o examples).