ldenoue/pulse-lfm25-1.2b-coach-onnx

Pulse LFM2.5 1.2B Coach ONNX

This is a coaching-focused fine-tune of LiquidAI/LFM2.5-1.2B-Instruct, packaged as a WebGPU INT4 ONNX model for in-browser inference with Transformers.js.

The model is designed for brief workplace-coaching conversations. It was trained to remain in the role of a company coach, use prior conversation turns, give concise spoken responses, and redirect requests that are outside a normal coaching session.

Modifications

• Fine-tuned with LoRA on 1,000 synthetic multi-turn coaching sessions, split into 900 training sessions and 100 evaluation sessions by scenario family.
• Trained with response-only supervision: full conversation history was visible as context, while only coach responses contributed to the loss.
• Selected the second-epoch checkpoint using the lowest held-out validation loss.
• Merged into the original LFM2.5 1.2B Instruct checkpoint.
• Exported to ONNX Runtime GenAI INT4 for WebGPU.
• Normalized external-data filenames and LFM convolution/KV-cache metadata for Transformers.js compatibility.

The training data was generated with Gemini and programmatically checked for role ordering, verbosity, Markdown, identity leaks, duplicate sessions, and conversation-history structure. It has not been exhaustively reviewed by human coaching professionals.

Browser usage

import { pipeline } from "@huggingface/transformers";

const coach = await pipeline(
  "text-generation",
  "ldenoue/pulse-lfm25-1.2b-coach-onnx",
  { device: "webgpu", dtype: "q4" },
);

The package uses onnx/model_q4.onnx and its adjacent external weight file. It was verified with Transformers.js 4.2.0 and ONNX Runtime Web 1.29.0.

Intended use and limitations

This model is intended for low-stakes workplace reflection, communication practice, prioritization, leadership, and professional growth. It is not a therapist, medical or legal professional, emergency service, HR decision-maker, or source of live factual information. Applications should retain explicit safety escalation rules and evaluate outputs before production use.

The fine-tune may reproduce biases or errors from the base model and synthetic dataset. Quantization can also change behavior. Do not use the model for employment decisions, diagnosis, crisis response, or other high-impact automated decisions.

License and attribution

Distributed under the LFM Open License v1.0. The included LICENSE is copied from the base model. Liquid AI created the original model; this repository contains a modified fine-tune and ONNX conversion by Laurent Denoue for the Pulse AI coach project. “Liquid AI” and “LFM” are used only to describe the model's origin.