TabRL

The question

Generating a table requires deciding its structure as well as filling its cells. Separating those tasks makes intermediate outputs inspectable.

How it works

• Generate a typed JSON schema before producing a schema-conditioned table.
• Prepare instruction data and train both LoRA stages using distributed training workflows.
• Retain malformed outputs and structured predictions for error analysis and inference validation.

Evidence & results

• Two fine-tuned checkpoints are publicly available.
• The final team report explicitly records the contribution split and training approach.
• Stored artifacts include 120 schema predictions and 120 table outputs for error analysis.

Scope & next steps

• Team implementation, not sole authorship.
• The reinforcement-learning extension is a design, not a demonstrated source of improvement.
• Reported prompting comparisons do not establish state-of-the-art quality or RL gains.