Alex — French/English front-desk voice assistant (LFM2.5-Audio-1.5B fine-tunes)
Speech-to-speech fine-tunes of LiquidAI/LFM2.5-Audio-1.5B for a company front desk: Alex welcomes visitors in French or English, finds the employee they came to see, notifies them on Teams, calls a receptionist when needed, gives practical information, and holds a conversation on any topic. Four tools, the LFM2 tool-calling protocol (a text-only decision turn, then the spoken turn).
Runs
| Folder | What it is |
|---|---|
runs/s2_r8/final |
R8: continued SFT from baptistefrancois1/lfm2a-fr@s12-unique-sft (all weights, lr 5e-5, one epoch, batch 64, context 2048) on long dialogues aimed at the failures measured on simulated visitors (Teams failure followed by the backup, a "no" kept in mind for the rest of the conversation, illness, unknown employees, refused names, relaunches, language switches, open conversation on any topic), with a full replay of S12's corpus so nothing is forgotten. |
runs/s2_r8b/final |
R8b: S12's single-phase recipe from the base model (lr 1e-4, one epoch) with the R8 dialogues added. |
Each folder holds model.safetensors (weights only), params.json (training settings) and, when available,
eval/*/summary.json (INCLUDE and Belebele log-likelihood text benchmarks, tool-call probes). The tokenizer, the Mimi
codec and the audio detokenizer are those of the base release (revision c362a0625dfe45aa588dce5f0ada28a7e5707628).
Use
Served with liquidai-vllm (vLLM-Omni, OpenAI-compatible API), which completes a weights-only folder from the base release:
huggingface-cli download Rcarvalo/lfm2a-fr-front-desk --include "runs/s2_r8/final/*" --local-dir alex
liquidai-vllm serve --model alex/runs/s2_r8/final --base-model LiquidAI/LFM2.5-Audio-1.5B
liquidai-vllm voice --url http://localhost:8000 # hands-free voice page
Decoding used in training and evaluation: audio temperature 0.7, top-k 4; text greedy. System prompt: "You are Alex, the voice assistant at Sia's front desk. You welcome visitors and use the tools available to you to help them. Reply in the language the user speaks." followed by the tool list.
Limits
Research checkpoints, evaluated on synthetic and simulated visitors and a small set of real voices. The directory, company and people in the training dialogues are fictional. Keep a human receptionist in the loop.