IFStruct v1.0
Chat & Writing · 2026-06-30
Liquid AI's IFStruct measures structured-output compliance: whether a model, without constrained decoding, emits valid JSON or YAML that satisfies a requested schema plus sampled item-count, wrapper-key, code-fencing and no-commentary constraints phrased in varied, often humanlike ways. Version 1.0 is a frozen public test set of 2,000 prompts scored pass/fail by a deterministic validator that also fails any unrequested key; content quality is not judged.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-1.7B | 93.7 |
| Qwen3.5-9B | 89.2 |
| Qwen3.5-4B | 87.0 |
| LFM2.5-2.6B | 82.2 |