Private tutorial data for notebooks/LFM3_Audio_(2B)-FR-tools.ipynb (rcarvalo/liquidai-vllm): web_search / db_query dialogues as LFM3-Audio messages. One row per exchange: system (routing rules + the model card's List of tools:), tools (JSON, apply_chat_template(tools=...)), messages (JSON, LFM3's own protocol: an assistant message with tool_calls, a tool message for the result, the reply as said), call (the expected decision, JSON or null) and reply (text in spoken form: numbers in words).

config / split	rows	fr	en
tools / train	3060	2447	613
tools / validation	65	53	12
spoken / train	4330	3424	906
spoken / validation	95	76	19

spoken/ holds the stage-3 rows: the same tool dialogues under the spoken system prompt (Respond with function calls, and interleaved text and audio.), plus social conversation windows answered without a tool (source = conversation, from Rcarvalo/conversational-s2s-v1). Their replies are voiced by the base model's own French and English TTS voices and stored as codec frames under spoken/codes/ (id, frames as a flat int16 list, n_frames, audio_s), the targets of the interleaved mode.

Sources (by the same author, gated): Rcarvalo/lfm25-fr-corpus-v1 (dataset/toolcalling_fr.jsonl, dataset/toolcalling_en.jsonl) and Rcarvalo/tc-en-voice-agent-v1 (phase_b/train_v5_1.jsonl). English is sampled to one row in five (seed 0). The 267 dialogues of the agent benches (reports/agent_eval/*.jsonl) are held out by id. The voice rows of stage 1 are Rcarvalo/lfm25-audio-fr-tutorial (SIWIS, CC BY 4.0).

Derived data for use with Liquid AI's LFM3-Audio-2B: LFM Open License v1.0. Not for publication without review.

Downloads last month: 30

Total file size:

8.16 MB