Surge AI Releases High-Quality Datasets: STEM Data Unexpectedly Boosts Coding by 6pp
echen · x · 2026-08-01
Surge AI's research team shared concrete performance gains on frontier models using their high-quality post-training datasets. Experiments show that training on just a fifth of their data yields significant improvements across various benchmarks:
- Coding: +12.0pp on Terminal-Bench 2.0
- Enterprise Agents: +9.6pp on Toolathlon
- STEM Reasoning: +9.3pp on FrontierScience
Interestingly, cross-domain generalization was observed: training on non-coding (e.g., STEM) datasets still improved coding performance by 6pp on SWE-Bench Pro without containing a single line of code. The full datasets are now available off-the-shelf with a 'pay-if-it-moves-your-metric' model.
More from Models
- 1-bit Quantization Shrinks Kimi K3 by 62% While Retaining 1M Context — rohanpaul_ai · 2026-08-01
- 1-bit Kimi K3 Quant Tested: 2.8T Model Compressed to 590GB Runs Locally — rohanpaul_ai · 2026-08-01
- Open Source Miniaturization: Opus-Level Models on Laptops Next Year? — No-Meringue5867 · 2026-08-01
- Testing DeepSeek v4 Flash: Major Improvements in UI/UX Design — curiousily_ · 2026-08-01
- Inkling-Small Local Test: Accurately Parses 85-Cell Medical Lab Reports on Mac — MaziyarPanahi · 2026-08-01
- Surge AI Releases High-Quality Datasets: Non-Coding Data Boosts Coding Benchmarks — echen · 2026-08-01