Papajams/orbura-dataviz-llama32-3b-autoscientist

Orbura AutoScientist Data Visualization Model (Llama 3.2-3B)

Model Details

• —Base model: meta-llama/Llama-3.2-3B-Instruct
• —Fine-tuning method: LoRA (r=16, alpha=32) via AutoScientist co-optimization
• —Dataset: Papajams/orbura-dataviz-augmented (5,000 rows with reasoning traces)
• —Task: Data visualization — matplotlib code generation, chart QA, code-to-description, style transfer, bug repair
• —Competition: AutoScientist Challenge Part 2 — Data Visualization

Training

• —Train examples: 5,000 (augmented with reasoning traces via Adaption Adaptive Data)
• —Validation examples: 2,000 (deterministic, held-out)
• —Iterations: 3 (AutoScientist co-optimization loop)
• —Best win rate: 52.01%
• —LoRA rank: 16, alpha: 32
• —Learning rate: 1e-5, cosine schedule
• —Epochs: 1
• —Training type: LoRA
• —Batch size: max (platform-selected)
• —Compute: Free via Adaption voucher

Evaluation

Naive baseline: 17.6% on 2k validation set.

Run the eval harness:

Usage

This is a LoRA adapter. Load it with PEFT:

License

Apache 2.0 (base model license applies). Dataset and model generated for the AutoScientist Challenge by Adaption Labs.