Train a lightweight Phi‑2 model using LoRA to perform chain‑of‑thought reasoning on 2D grid maps. Solve simple maze or spatial planning problems step by step.

• Grid‑based visual reasoning using spatial CoT prompts
• LoRA fine‑tuning to keep compute and model size low
• Example maze puzzles with Web UI (map_interface.html)
• Training and evaluation scripts included (plot_trainer_state_cli.py, flask_api.py)

Simulate an agent navigating a 10x10 grid using discrete action steps.
The objective is to compare different input formats and reasoning strategies:

• CoT with vector inputs
• NLP-based commands
• Direct vector-to-position reasoning (baseline)

Each model is under outputs/, and each .bin file is under 100MB.

• 11x11: basic spatial trace task, vector action → position (no map)
• 11x11 map: adds map context to input, model must parse visual structure
• label: data includes correctness classification (correct, loop, etc.)

GPT-CoT/
├── configs/              # LoRA training config files (YAML)
├── data/                 # JSONL training files
├── outputs/              # Fine-tuned models (3 total)
│   ├── phi2-CoT-finetune5/
│   ├── phi2-CoT-finetune11x11/
│   ├── phi2-CoT-finetune11x11_map/
│   ├── phi2-Label-finetune1/
│   ├── phi2-NLP-finetune1/
│   └── phi2-vec-finetune/
├── source/               # Training and inference scripts
├── .gitignore
├── README.md
└── requirements.txt

git clone https://github.com/Seanaaa0/GPT-CoT.git
cd GPT-CoT
conda activate gpt-env  # or your preferred environment
pip install -r requirements.txt

• New fine-tuned model: phi2-Label-finetune1
• Task: Given a series of vector actions (dx,dy), reason step-by-step to compute the final position and classify the path as one of:
• correct, too short, too long, loop, out of bound, wrong
• Training data: 10x10_vec_labeled.jsonl
• Inference script: inference_phi2_vec.py
• Accuracy: ~95%, supports full CoT + label correctness tracking
• Output example includes "label" and "correct" field for each prediction

• map_interface.html: displays a 10x10 grid and agent paths interactively
• flask_api.py: serves model predictions and links frontend ↔ backend
• Future integration with live inference and editing

We provide a Python tool to visualize inference traces from test_label.jsonl.

This script will:

• Parse GPT output traces
• Generate per-sample visualizations
• Combine up to 25 images into a grid

Use plot_trainer_state_cli.py to visualize training loss and gradients:

python plot_trainer_state_cli.py --file results/trainer_state/trainer_stateX.json --metrics 1

• Metric options:
  • 1: Loss
  • 2: Grad norm
  • 3: Learning rate

Output PNG files are saved to results/png/.

cd source/data/test_output
python generate_trace_images.py

---

##  TODO
- [✅] Train LoRA on vector trace task
- [✅] NLP command version
- [✅] Multi-entry point generalization
- [✅] Trace classification (valid/invalid)
- [ ] Decision Transformer for path generation
- [ ] Add goal-aware discriminator

---

## 📜 License
MIT