learosema's picture
Upload README.md with huggingface_hub
b5c9ef7 verified
metadata
language: en
license: apache-2.0
tags:
  - webbee
  - code-generation
  - text2text-generation
  - t5
  - onnx
datasets:
  - synthetic
pipeline_tag: text2text-generation

WebBee Delegate Models

Fine-tuned T5-small models used by WebBee — a plain-English web development agent — for its Tier 2a inference pipeline.

Each model is an ONNX int8-quantised export of a T5-small checkpoint trained on synthetic data. They run locally via @xenova/transformers with no GPU required.

Models

goal-to-steps/

Decomposes a high-level development goal into a sequence of concrete WebBee commands. This is WebBee-specific: the model knows the agent's verb vocabulary (add, inject, install, add route, add hook, …) and the idiomatic step format used throughout the system.

Task text2text-generation
Input Natural-language goal (e.g. "add authentication")
Output Pipe-separated WebBee steps (e.g. "add component LoginForm | add hook useAuth | add route api/auth")
Training examples 533 synthetic goal→steps pairs across 8 categories

Usage

import { pipeline } from '@xenova/transformers';

const generator = await pipeline(
  'text2text-generation',
  'learosema/webbee-delegate-models',
  { subfolder: 'goal-to-steps' }
);

const result = await generator('add authentication');
console.log(result[0].generated_text);
// → add component LoginForm | add hook useAuth | add route api/auth

Training

Models were trained and exported using the scripts in packages/delegate-model/ of the WebBee repository.

Base model: google/t5-small
Quantization: ONNX int8 via optimum

License

Apache 2.0 — same as the base T5-small model.