---
license: apache-2.0
tags:
- evaluation
- speech
- traces
---

# FlexiSLM model evaluation traces

JSONL inference traces only (no generated WAVs).

Each file is named after the model. Records follow the FlexiSLM eval trace schema
(`index`, `task`, `input`, `output`, `evaluation`, `metadata`).

## Files

- `flexislm_stage2_released.jsonl` — FlexiSLM-7B Stage2 released checkpoint (ckpt-150000), 12.5 Hz in/out, flow-matching s2s on VoiceBench + OpenAudioBench.
- `flexislm_stage3_ckpt15000.jsonl` — FlexiSLM-7B Stage3 moreframesv2 checkpoint-15000, 12.5 Hz in/out, flow-matching s2s on VoiceBench + OpenAudioBench + LibriSpeech ASR (test-clean/other).
- `flexislm_stage3_ckpt20000.jsonl` — FlexiSLM-7B Stage3 moreframesv2 checkpoint-20000, same benchmark suite as ckpt-15000.