Laguna XS.2: Synthetic Training Data for Coding and Scientific Reasoning

Synthetic training data generated in May 2026 as part of the Poolside Research Hackathon. It is intended to improve the Laguna-XS.2 model on coding and scientific reasoning tasks. The data was produced by the author 'poolside-laguna-hackathon' using the Qwen3.6-35B-A3B model on the SWE-bench source dataset.

Use Cases

• Supervised fine-tuning (SFT) for code generation models based on the provided teacher solutions and problem statements.
• GRPO training to improve model performance on coding tasks using the structured trajectories and scores.
• Benchmarking model improvements on SWE-bench problems using the synthetic trajectories.
• Research into synthetic data generation methods for programming and scientific reasoning.

Strengths

• Data is specifically generated for a defined purpose: improving the Laguna-XS.2 model on coding and scientific reasoning.
• Source dataset is explicitly named as SWE-bench, a known benchmark.
• Teacher model used for generation is specified as Qwen3.6-35B-A3B.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Synthetic data generated from the SWE-bench dataset using the Qwen3.6-35B-A3B model.

Collection Method

Produced as part of the Poolside Research Hackathon (May 2026).

Time Range

Generated in May 2026.

Freshness

Last updated 2026-05-30 11:49:03; freshness should be verified.