File size: 7,365 Bytes
36f82a2 2b2082a 36f82a2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 | ---
license: apache-2.0
---
# TensorTalk / UM_Handbook
TensorTalk is a handbook-grounded academic chat assistant built for the **Faculty of Computer Science and Information Technology, Universiti Malaya (UM)**.
This project focuses on turning UM handbook content into a usable question-answering system through:
- handbook preprocessing
- source chunk construction
- supervised QA dataset building
- Qwen3-8B LoRA fine-tuning
- merged-model deployment
- a browser-style HTML chat demo
---
## Project Goal
The main goal of this project is to build a handbook-based assistant that can answer student questions using information learned from the UM handbook domain.
The current version is designed around:
- undergraduate and postgraduate handbook content
- handbook-faithful answers
- concise student-facing responses
- a local/demo deployment workflow on DICC and notebook environments
This project is also intended to support a broader experimental pipeline:
- **Baseline 1:** closed-book supervised fine-tuning
- **Baseline 2:** retrieval-augmented version for later comparison
---
## What This Project Contains
### 1. Dataset Preparation
The project includes scripts and resources for preparing handbook data before fine-tuning:
- handbook markdown preprocessing
- source chunk dataset building
- SFT QA dataset construction
- configuration management for the preprocessing and dataset pipeline
### 2. Fine-Tuning Workflow
The model training workflow uses a Qwen3-8B base model with LoRA-based fine-tuning on the UM handbook QA dataset.
The fine-tuning workflow includes:
- notebook-based training on DICC
- device-aware loading logic
- train / validation / test style evaluation workflow
- merged-model export for direct inference
- LoRA adapter export for optional PEFT-based reuse
- metrics and prediction file generation
### 3. Deployment Demo
The project includes a notebook-based HTML chat UI called **TensorTalk**.
The demo provides:
- a browser-style chat layout
- a handbook-focused system prompt
- merged-model loading for direct inference
- a student-facing question-answer workflow
- a simple deployment path for demonstration purposes
---
## Current Project Structure
```text
UM_Handbook/
├── Dataset/
│ └── SFT_Dataset/
│ ├── SFT_QA_Training_Ready.jsonl
│ ├── SFT_QA_Training_Ready_pretty.json
│ ├── SFT_QA_Metadata.jsonl
│ └── SFT_QA_Metadata_pretty.json
├── assets/
├── outputs/
│ └── qwen3_um_handbook_optimized_1/
│ ├── lora_adapter/
│ ├── merged_model/
│ ├── trainer_runs/
│ ├── test_eval_runs/
│ ├── dataset_split_summary.json
│ ├── final_metrics.json
│ ├── test_predictions.jsonl
│ └── validation_predictions.jsonl
├── FineTune_QWEN3_UM_Handbook_optimized_1.ipynb
├── UM_Handbook_Markdown_Preprocess.py
├── UM_SFT_QA_Dataset_Builder_from_Index.py
├── UM_Source_Chunk_Dataset_Builder.py
└── um_handbook_config.py
```
---
## Key Files
### Training and Data
- `Dataset/SFT_Dataset/SFT_QA_Training_Ready.jsonl`
Main SFT training dataset used for handbook QA fine-tuning.
- `UM_Handbook_Markdown_Preprocess.py`
Preprocesses handbook markdown / extracted source text.
- `UM_Source_Chunk_Dataset_Builder.py`
Builds source chunks for downstream dataset and retrieval-related use.
- `UM_SFT_QA_Dataset_Builder_from_Index.py`
Builds the supervised QA dataset from curated handbook content.
- `um_handbook_config.py`
Central configuration file for paths and data-processing settings.
### Training Output
- `outputs/qwen3_um_handbook_optimized_1/merged_model/`
Main inference-ready model directory.
This is the directory used by the demo chat UI.
- `outputs/qwen3_um_handbook_optimized_1/lora_adapter/`
LoRA adapter weights.
This is useful for PEFT-style loading with a base model, but it is not the primary path used by the current demo UI.
- `outputs/qwen3_um_handbook_optimized_1/final_metrics.json`
Final evaluation summary.
- `outputs/qwen3_um_handbook_optimized_1/validation_predictions.jsonl`
Validation-set generated answers for inspection.
- `outputs/qwen3_um_handbook_optimized_1/test_predictions.jsonl`
Test-set generated answers for inspection.
### Demo
- `FineTune_QWEN3_UM_Handbook_optimized_1.ipynb`
Main notebook that contains the fine-tuning workflow and the TensorTalk HTML chat demo.
---
## Model Artifact Notes
This project may contain several model-related outputs. They are not all used in the same way.
### `merged_model/`
This is the most important deployment artifact for the current demo.
Use this when:
- running the current TensorTalk HTML chat UI
- loading the fine-tuned model directly with Hugging Face `from_pretrained(...)`
- sharing the main inference-ready model
### `lora_adapter/`
This contains LoRA delta weights only.
Use this when:
- loading the adapter on top of the original base model
- reusing the fine-tuning result in a PEFT workflow
- experimenting with a smaller transferable fine-tuning artifact
### `.pt` exported model file
If present, the `.pt` file is mainly a saved full-model artifact / backup export.
Use this when:
- archiving the full fine-tuned weights
- running a custom loading workflow that explicitly expects a `.pt` file
For the current TensorTalk chat UI, the primary runtime artifact is still **`merged_model/`**.
---
## Current Demo Behavior
The current demo is designed to answer questions such as:
- dress code and appearance guidance
- programme core courses / credit requirements
- undergraduate vs postgraduate handbook information
- academic rules and handbook-supported policy questions
The answer style is intended to be:
- handbook-grounded
- short and direct
- student-facing
- non-speculative
---
## Example Demo Output
The screenshot below shows the current TensorTalk chat interface running with the fine-tuned UM handbook model.
---
## Repository Preview
The screenshot below shows the current top-level project layout.
---
## Suggested Minimal Deployment Package
If the goal is only to demonstrate the chat UI to teammates, the minimal useful set is:
- `merged_model/`
- the chat notebook / UI code
- optional avatar image under `assets/`
The following items are not required for a simple demo run:
- intermediate training checkpoints
- test evaluation run directories
- optional full `.pt` export
- raw training logs not used by the demo
---
## Notes
- The project is organized so that **Dataset**, **models / outputs**, and **demo code** remain separate.
- The current demo is notebook-friendly and was prepared around a DICC workflow.
- The deployment path prioritizes clarity and reproducibility over a heavyweight full-stack application setup.
---
## Status
Current project status:
- handbook preprocessing pipeline prepared
- supervised QA dataset prepared
- LoRA fine-tuning workflow completed
- merged model exported
- TensorTalk HTML chat demo running
- evaluation outputs generated
---
## Author / Project Name
**TensorTalk**
UM Handbook QA / Fine-Tuned Qwen3-8B LoRA Project |