Skip to content
← Models
Open weightsZhejiang University/10 trustVerified

TableGPT2

Qwen2.5-based 7B model with a dedicated table encoder for NL-to-SQL, BI analytics and messy real-world spreadsheets.

Visit official page →

About

TableGPT2 is a 7B model built on Qwen2.5 that adds a dedicated table encoder to the standard decoder, designed to read schemas and cells the way a vision encoder reads pixels. It was trained on 593,800 tables and 2.36M query-table-output pairs, and it is explicitly aimed at the messy reality of business data: ambiguous questions, missing column names and irregular spreadsheets. The model card notes a Chinese-heavy corpus, so non-Chinese queries may be weaker.

On the authors' own benchmark tables it leads open 7B-class models on NL-to-SQL (Spider test 74.38, BIRD dev 31.42 execution accuracy) and on the holistic TableBench suite, while keeping its general coding and language ability intact. The team also ships tablegpt-agent on GitHub for agentic BI workflows. The weights are Apache 2.0 and public on Hugging Face. The main caveat is that the headline numbers are author-reported with no independent replication yet, and the 72B variant in the paper has not been released.

Specs

Context window

32,768 tokens (Qwen2.5-7B base)

Pricing

Free / open weights

Free quota

Weights free for research and commercial use under Apache 2.0

API access

Hugging Face (tablegpt/TableGPT2-7B); tablegpt-agent; self-hosted

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Spider test execution accuracy: 74.38 (7B) (author-reported, arXiv 2411.02059 Table 5)
  • BIRD dev execution accuracy: 31.42 (7B); 60.76 with external knowledge (author-reported)
  • TableBench DP / TCoT / SCoT / PoT@1: 32.03 / 42.34 / 25.01 / 33.52 (7B) (author-reported)
  • Average +35.20% (7B) / +49.32% (72B) over prior benchmark-neutral LLMs across 23 metrics (author-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-sqldata-analysisbusiness-intelligencespreadsheets

Trust score — 7/10

/10 trustVerified

Scored 7/10 (TIER metrics, metrics-v1). Evidence: arXiv 2411.02059 (Nov 2024); HF tablegpt/TableGPT2-7B with full model card (developers: Zhejiang University); paper Table 5 reports Spider test 74.38 / BIRD dev 31.42 for the 7B; companion tablegpt-agent repo. Strengths: dedicated table encoder, strong NL-to-SQL results, Apache 2.0 weights, retained general ability. Docked for: (1) all benchmark numbers are author-reported, no independent replication found; (2) 72B variant not released; (3) Chinese-heavy corpus, weaker on other languages. License: Apache 2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · arXiv 2411.02059 (v1 2024-11-04); developers Zhejiang University; HF tablegpt/TableGPT2-7B with full model card
  • · Trained on 593.8K tables + 2.36M query-table-output tuples (paper)
  • · Paper Table 5: Spider test 74.38 / BIRD dev 31.42 exec accuracy (7B); +35.20% avg over prior LLMs on 23 metrics (author-reported)
  • · Companion tablegpt-agent repo for agentic BI workflows
  • · Benchmarks author-reported, no independent replication; 72B variant not released; Chinese-heavy corpus

Sources