📝 Blog Mar 8
Format Compliance as Separate Capability: Small Models Lack It
Production testing reveals Gemma 12B and Qwen 3.5 35B return correct answers in unparseable formats despite explicit instructions—Python instead of CSV, Markdown instead of CSV. Format compliance is independent capability missing from all major benchmarks (SWE-bench, Aider, LiveBench, SEAL), critical gap for production pipelines where consumers are parsers not humans. Smaller models fundamentally lack instruction-following precision for machine-readable output.