TR4F/crackio-talk

Crackio Talk voice

A LoRA on Liquid AI's LFM3-VL-2B that makes it talk like a person on a voice call: short, warm, casual replies, the right length for what was said (a few words for a few words), with no lists or formatting. Crackio's Talk mode hears you, has LFM3-VL-2B answer with this LoRA and speaks each sentence as soon as it's written, all on your own GPU.

Files

File	What it is
lfm3-vl-2b-lora-chat.safetensors	The LoRA as Crackio's ComfyUI bridge loads it, from ComfyUI/models/lfm3/
peft/	The same LoRA as a PEFT adapter (rank 32, alpha 32, on the language model's attention and MLP)

Use it

In Crackio, setting up Voice in Settings → Tools with Talk's chat voice puts LFM3-VL-2B and this LoRA on your GPU, and Talk uses it automatically. Crackio's bridge serves it at POST /crackio/lfm3 (or /crackio/lfm3/stream) with "lora": "chat" and the conversation so far as history.

With PEFT, load LiquidAI/LFM3-VL-2B and apply peft/ on top. It was trained with this system prompt:

You're talking out loud with someone. Sound like a real person: short, warm and natural, the right length for what they said (a few words for a few words). No lists, no emojis, no formatting.

How it was trained

Three epochs on 89 short text conversations, rated on Crackio's Rate page. Only the last reply of each conversation was learned. The turns before it were context.

• The replies it learned. 66 were written by the Crackio author in their own words. The other 23 the author picked from two answers LFM3-VL-2B itself gave, each written under one of 14 short-speech instructions. No other AI model's replies were used.
• The other side's lines. These are 89 everyday moments such as "nice", "tell me a joke" or "my dog is sick", 17 of them a few turns long. Claude (Anthropic), Crackio's coding assistant, wrote them for Crackio's tests. They were only prompts. The earlier replies inside the longer conversations are LFM3-VL-2B's own.

Results. These are from the first build, which had 8 more moments with a picture. On 18 held-out moments its replies averaged 13.8 words, against 17.3 for LFM3-VL-2B alone (medians 10.5 and 14), and they read more casually.

Known limits.

• Some questions ("do you like music?") still get long, assistant-like answers, and a few replies are odd.
• It learned from 89 examples, so it may repeat some of the author's replies almost word for word.
• It was trained on English conversations only.

Licence

LFM3-VL-2B and this adapter, a Derivative Work of it, are under the LFM Open License v1.0 (see LICENSE). Commercial use is licensed only for companies with under US$10M annual revenue. See NOTICE for attribution and what was changed.