Tessar (Large-Sized Model)

Tessar is an advanced table reasoning model developed by SVECTOR, building upon the groundbreaking research and pushing the boundaries of neural table understanding.

Model Description

Tessar (Textual SQL Analysis and Reasoning) is a sophisticated neural model designed to excel in table-based question answering. Tessar implements an innovative neural SQL executor that can interpret and reason over complex tabular data with remarkable precision.

The model is constructed using the BART architecture, featuring a bidirectional encoder and an autoregressive decoder. This design allows Tessar to capture intricate contextual relationships within tabular data and generate accurate, contextually relevant answers.

Key Features

Advanced neural SQL execution capabilities
State-of-the-art performance on complex table question answering
Robust handling of nuanced and multi-step queries
Fine-tuned on the WikiTableQuestions dataset

Intended Uses

Tessar is particularly powerful for solving complex table-based questions across various domains. Here are some example questions the model can effectively address:

Question	Example Answer
According to the table, what is the last title produced?	Specific Title
What is the difference in a specific comparative metric?	Numerical Difference
Which entity had the most significant impact in a given context?	Identified Entity
What were the first and last entries in a specific column?	Comparative Entries

How to Use

Here's a comprehensive example of using Tessar with the Transformers library:

from transformers import TessarTokenizer, BartForConditionalGeneration
import pandas as pd

# Load Tessar model and tokenizer
tokenizer = TessarTokenizer.from_pretrained("SVECTOR-CORPORATION/Tessar-largest")
model = BartForConditionalGeneration.from_pretrained("SVECTOR-CORPORATION/Tessar-largest")

# Prepare sample table data
data = {
    "year": [1896, 1900, 1904, 2004, 2008, 2012],
    "city": ["athens", "paris", "st. louis", "athens", "beijing", "london"]
}
table = pd.DataFrame.from_dict(data)

# Ask a specific query
query = "In which year did beijing host the Olympic Games?"
encoding = tokenizer(table=table, query=query, return_tensors="pt")

# Generate answer
outputs = model.generate(**encoding)

# Decode and print result
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
# Expected output: [' 2008.0']

Evaluation

For comprehensive evaluation scripts and benchmarking, please refer to the SVECTOR documentation and research repositories.

Performance Highlights

Exceptional accuracy on complex table reasoning tasks
Robust handling of multi-step and contextual queries
State-of-the-art performance on WikiTableQuestions dataset

Citation

If you use Tessar in your research the SVECTOR implementation:

@inproceedings{svector2025tessar,
    title={{Tessar}: Advanced Neural Table Reasoning},
    author={{SVECTOR Team}},
    year={2025},
    institution={SVECTOR Research}
}

Contact and Support

For further information, support, or collaboration opportunities, please contact SVECTOR's research team at [email protected].

License

This model is released under the MIT License. Please review the licensing terms before use.

SVECTOR-CORPORATION
/

Tessar-largest