Human Instructions
On Human Instructions, AsyncIO performed worse and slower than the synchronous SFT baseline. The Human Instructions results suggest synthetic streaming training does not fully cover spontaneous spoken interaction. Naturalistic speech trans…
1 sources - 5 claims
On Human Instructions, AsyncIO performed worse and slower than the synchronous SFT baseline. The Human Instructions results suggest synthetic streaming training does not fully cover spontaneous spoken interaction. Naturalistic speech transcripts caused lower accuracy and increased AsyncIO latency due to repeated-action behavior. The final Human Instructions set contains 177 valid sessions after human review. The Human Instructions evaluation set targets natural speech phenomena missing from standard benchmarks.