It is designed for client-side execution with Transformers.js/WebGPU and a separate browser-side action validator. The training data uses curated public portfolio summaries and synthetic interaction examples; it is not intended to expose raw CV/PDF text, private records, schemas, patient data, supplier data, or customer data.

Files

onnx/model_q4.onnx
onnx/model_q4.onnx_data
tokenizer/config files required by Transformers.js
strict_eval_q4_500.json

Validation

The exported Q4 ONNX artifact passed the held-out strict browser-agent eval:

full_json: 1.0
json_object: 1.0
strict_valid: 1.0
action_exact: 1.0
citation_exact: 1.0
count: 500

Browser Usage

import { AutoModelForCausalLM, AutoTokenizer } from "@huggingface/transformers";

const modelId = "YMRohit/rohit-portfolio-lfm25-350m-onnx";
const tokenizer = await AutoTokenizer.from_pretrained(modelId);
const model = await AutoModelForCausalLM.from_pretrained(modelId, {
  device: "webgpu",
  dtype: "q4",
});