Rcarvalo/lfm25-agent-230m-train

LFM2.5-Audio 230M agent: training rows (web_search / db_query, French and English)

Training set of the 230M LFM2.5-Audio student distilled from lfm2a-fr S12, trained to decide when to call web_search (public, current facts) or db_query (internal company data), to speak about the result, and to keep a conversation. Built by scripts/build_agent_dataset.py of the unsloth-lfm2.5-audio tutorial (22/10/2026).

split	rows
train	21 556
validation	826

Train rows by source: French tool corpus 2 447, English v5.1 2 634, English corpus 1 940, generated French 4 298, generated English 2 791, social conversations 2 446, SIWIS replay 5 000 (ASR and TTS, half each).

Format. One row per training window, in the serving layout of liquidai-vllm serve --backend llama: system prompt with the two tools, earlier turns as text, the current user turn as audio (16 kHz FLAC in audio_in) or text, then the assistant turns (call, tool result, empty decision, spoken reply). Columns: id, lang, source, task, messages (JSON parts text / audio_in / interleaved / audio_out), audio_in, codes (the reply's Mimi codes, 8 codebooks x frames, end-of-audio frame included), text_kd (whether the teacher's text distribution is a target: off on tool rows, which the teacher never learnt). A window ending ~decision had no voiced reply and keeps its decision turns only.

Voices. Assistant: Qwen3-TTS fine-tuned on SIWIS (Rcarvalo/qwenTTS, French) and Qwen3-TTS CustomVoice "Serena" (English); users: Qwen3-TTS VoiceDesign, one designed voice per dialogue (gender, age, accent, pace), and the original user clips of the source corpora. Measured with Whisper large-v3-turbo, numbers normalized: French assistant 3.5 % WER, English assistant 7.2 %, visitors 5.4 %.

Text. Generated dialogues: Claude Sonnet 5.5, ten families (public and internal routing, ambiguous routing, follow-ups, grounded answers, absent answers, no call, clarification, chat then tool, reuse of a result), checked by a parser and the tool registry, filtered against the held-out test sets (written separately by Claude Opus 5.5). Conversations: social intents of Rcarvalo/conversational-s2s-v1.

License. Non-commercial: the French and English tool corpora (Rcarvalo/lfm25-fr-corpus-v1) are CC-BY-NC. SIWIS is CC-BY 4.0.