DePlot
Plot-to-table translation: turn a chart image into a linearized data table, then let an LLM do the reasoning — one-shot chart QA without fine-tuning.
About
DePlot decomposes chart question answering into two steps: first translate the chart image into a linearized table, then hand that table to a large language model and let few-shot prompting do the reasoning. The modality conversion module — an end-to-end image-to-text Transformer built on the Pix2Struct architecture — is trained specifically for plot-to-table translation on synthetic and web-crawled charts paired with their underlying data.
The paper (arXiv 2212.10505, from Google DeepMind with the University of Cambridge) reports that DePlot plus an LLM with one-shot prompting beat a SOTA model fine-tuned on 28,000+ examples by 24% on human-written chart QA queries. It is the cleanest demonstration that for charts, separating perception from reasoning beats end-to-end training.
Weights are on Hugging Face as google/deplot under Apache 2.0, and the code lives in the googleresearch repo. It is a specialist, not a generalist: for charts and plots it is excellent, for anything else look elsewhere.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2212.10505): DePlot+LLM with one-shot prompting achieves 24.0% improvement over a SOTA model fine-tuned on 28k+ examples, on human-written chart QA queries (provider-claimed)
- Built on the Pix2Struct architecture; code in googleresearch/deplot
- No independent replication of the headline number published
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: paper from Google DeepMind and Cambridge (arXiv 2212.10505) with a clear decomposition methodology, Apache 2.0 weights on Hugging Face (google/deplot), code in googleresearch, and native Transformers support. Strengths: fully open, elegant perception/reasoning split, strong author-reported chart QA results. Docked for: (1) headline number is author-reported with no independent replication; (2) narrow specialist scope and archival maintenance. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Paper from Google DeepMind and the University of Cambridge (arXiv 2212.10505, released 2022-12-20); plot-to-table translation built on the Pix2Struct architecture.
- · Weights at huggingface.co/google/deplot (Apache 2.0), code in googleresearch/deplot, native Transformers support.
- · Author-reported 24.0% gain over a 28k-example fine-tuned SOTA on human-written chart QA queries; no independent replication published. Narrow specialist scope; maintenance archival.