How juctxy/debate works
This repository contains a Gradio debate-argument evaluator backed by a feature-based classifier, plus a separate TinyLlama fine-tuning and inference workflow. The main user path is entering debate arguments in the app and receiving an evaluation; the sampled excerpts show the interface and analyzer, but not the complete callback-to-result path, so that connection is omitted. The training script prepares debate prompts, fine-tunes a model, and demonstrates text generation. Some runtime details and connections to the included model assets are not shown.
Main entry points: full_step.py Prompt preparation · adapter_config.json TinyLlama adapter · app.py Argument analyzer · model.joblib Classifier assets
Components 10
Debate evaluator
- Debate interface
app.py - Argument analyzer
app.py - Winner classifier
app.py
LLM workflow
- Debate dataset
- Prompt preparation
full_step.py - Model fine-tuning
full_step.py - Text generation
full_step.py
Model assets
- Classifier assets
debate_2/model.joblib - TinyLlama adapter
TinyLlama-1.1B-chat-finetuned-debate/adapter_config.json
Other components
- Debater
Connections 6
- Debater → Debate interface: enters arguments
- Prompt preparation → Debate dataset: reads CSV
full_step.py - Prompt preparation → Model fine-tuning: provides dataset
full_step.py - Model fine-tuning → TinyLlama adapter: saves adapter
full_step.py - Text generation → TinyLlama adapter: loads adapter
- Argument analyzer → Classifier assets: loads components
app.py