One model that scores how much a chart lies, submitted to all seven Part 2 categories. A chart can be true in every value and still misrepresent its result. PolyChart measures how much a chart distorts its data, giving every distortion an exact lie factor, signed and continuous, rather than a yes/no label. Everything below is open: datasets and weights on HuggingFace and Kaggle, a Kaggle writeup and two notebooks, live demos, the client that drove the whole run, and the research it produced.
Since September 2026 I have been a volunteer ambassador for Adaption Labs — running events and community initiatives, testing new formats and platform features, and carrying member feedback back to the team.
Announced 18 August 2026. Marks and palette are Adaption Labs' own.
$4,0001st place in Data & Visualizationannounced 18 August 2026
“A visualization-reasoning model designed to determine whether claims made from charts are actually supported by the underlying data and to detect misleading visual patterns.”
Part 2 ran seven categories — Science, Agriculture, Personal Finance, Data & Visualization, Market Analysis & News, HR, and Math & Code — with a first and a runner-up named in each. The runner-up in this category was Vinod Anbalagan. Data Visualization is the entry that placed, and the other six shared its model and thesis. All seven are under Entries.
How the 1st-place Data & Visualization entry at the AutoScientist Challenge 2026 was built: exact lie-factor labels, more than 200 fine-tunes, and what the scores can and can't tell you.
Entries
One idea, tested seven ways. Each category is its own dataset and training run, evaluated on the platform's held-out test set for that category. All seven share one LoRA, linked under Code. Submissions ran 6 July – 10 August 2026.
Data Visualization
1st place
85.9held-out
A model that scores how much a chart lies. Every distortion carries an exact lie factor, signed and continuous, not a yes/no label. Ships with a Deception Atlas of 20 real published misleading charts, each checked against the sources recorded with it.
Scientific claims travel as charts, and a chart can be true in every value and still misrepresent its result. This checks the claim against the plotted data, not the picture.
News charts are the most published and least audited charts there are. This pairs chart reading with severity scoring, so a claim in a story can be checked against the figure that supposedly supports it.
Workforce reporting is full of traps: a comparison across differently defined cohorts, an axis choice that carries a conclusion the data does not. It reads participation and trend claims against the underlying numbers.
Quantitative reasoning over tabular and plotted data. Read exact values, compute ratios and rates of change, and refuse to answer when the number is not in the input. The refusal case matters as much as the arithmetic.
Financial charts are where misleading encodings do the most direct damage: a truncated axis on a returns chart, a percentage gain quoted without its base. It separates the visual impression from the arithmetic and reports the size of the gap.
Agricultural yield and input data, where long series and unit changes make trend claims easy to overstate. Measure first, then claim only what the data supports.
The parts you can use: the shared weights, two Kaggle notebooks that open the training data and the adapter, and the open-source client that ran every job.
Model weights
Shared model weights
A single LoRA, PolyChart — "Shown Is Not Supported", released for every category on both platforms.
To run at the scale this needed (more than 200 fine-tuning jobs), I built a headless client for the AutoScientist API before the official one shipped, and drove everything through it rather than the interface. It is what made the paired within-seed design and the run count feasible. I have open-sourced it so anyone can build on it.
Every dataset and notebook from this build, with the rest of my Kaggle work, is on my Kaggle page.
Papers
Two papers came out of this build. Both are preprints on Zenodo, and both are submitted for peer review. The fuller list lives on my research page.
Submitted to the International Journal of Data Science and Analytics · preprint on Zenodo
A Hosted LLM Fine-Tuning Service Kept Most Ordinary Chart Question-Answering Targets but Replaced Many Dataset-Specific Definitions and Abstentions: A User-Side Case Study With Answer Keys and Replicates
The measurement paper from this build. It audits the platform from its own per-job records for all 384 fine-tuning jobs on 21 datasets, driven headlessly through its internal API. The core capability works: adding 2,000 synthetic domain rows raised the held-out win rate on 10 of 10 paired seeds (mean +13.4, 95% bootstrap CI 11.7 to 15.1, exact sign test p=0.0020), with nothing measurable added beyond 2,000 rows. The numbers a user sees by default mislead: identical runs vary with a pooled standard deviation of 2.75 points, best-of-n reporting adds 2.1 to 6.3 points, and the displayed win rate orders 44.2% of same-dataset job pairs differently from the held-out rate. Every statistic was recomputed from the per-job records in October 2026, correcting nine statements of the August preprint. The revised version (submitted to the International Journal of Data Science and Analytics in October 2026, after Harvard Data Science Review did not invite a proposal) adds 97 jobs on a second base model across 16 other held-out domains and checks the platform's rewritten training data against answer keys it did not produce: the rewriting kept most ordinary answers correct but turned 8.2% to 19.6% of correct training targets into incorrect ones, mostly dataset-specific definitions and abstentions.
384 fine-tunes · +13.4 on 10/10 seeds (CI 11.7–15.1, p=0.0020) · corrected Oct 2026Preprint on Zenodo
Submitted to PacificVis 2027 · dataset released
Lie Factors by Construction: Exact Scale-Ratio Labels for Parameterised Chart Distortions, and an Inverted Axis That Shifts Seven of Nine Models' Direction Readings
The model paper behind PolyChart. Charts can mislead without a single false number: a truncated axis or a mismatched area encoding changes what a reader concludes while every label stays accurate. It builds a counterfactually-augmented dataset directly from the distortion parameters used to render each chart, so the ground truth for deception is generated rather than annotated — 2,430 rows on a 33-field machine-checkable schema, 636 carrying an exact lie factor — plus a held-out Deception Atlas of 20 real published misleading charts, each checked against the sources recorded with it. Training a continuous severity estimator with reinforcement learning from verifiable rewards is pre-committed in an analysis plan fixed before any run.