DrThibodeau — Alpha Graph Chat

This repository contains an experimental graph-guided Alpha language model and an interactive chat program.

Model files

• latest-inference-fp16.pt — inference-only FP16 checkpoint
• tokenizer.json — matching 32,000-token tokenizer
• graph_guided_alpha.py — model definition
• chat_graph_alpha.py — recommended chat interface

The checkpoint is approximately 63 million parameters with a 512-token context window. It is an experimental model and may produce incoherent or empty responses, especially after limited training.

Install

git clone https://huggingface.co/MaxiMThi/DrThibodeau
cd DrThibodeau

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

If the repository was downloaded without Git:

python -m pip install -U huggingface_hub
huggingface-cli download MaxiMThi/DrThibodeau --local-dir .

Chat

Run the recommended graph-enabled chat:

source .venv/bin/activate

python chat_graph_alpha.py \
  --checkpoint latest-inference-fp16.pt \
  --max-new-tokens 128 \
  --temperature 0.8 \
  --top-k 40 \
  --repetition-penalty 1.05 \
  --context-turns 4 \
  --prompt-format auto

On a CUDA GPU, the script uses CUDA automatically. On CPU, generation will be slower:

python chat_graph_alpha.py \
  --checkpoint latest-inference-fp16.pt \
  --max-new-tokens 64 \
  --temperature 0.7 \
  --top-k 20

Interactive commands:

/reset   Clear conversation history
/quit    Exit the chat

Generation controls

• Lower --temperature such as 0.3 makes output more deterministic.
• Higher --temperature such as 1.0 increases variation.
• Lower --top-k reduces unusual token choices.
• --context-turns controls how many previous turns are included.
• --prompt-format plain is useful for testing ordinary text completion.
• --disable-graph bypasses graph routing for an ablation comparison.

Example deterministic run:

python chat_graph_alpha.py \
  --checkpoint latest-inference-fp16.pt \
  --max-new-tokens 128 \
  --temperature 0.3 \
  --top-k 10 \
  --prompt-format auto

Important compatibility rule

Keep these files together and use the graph chat script:

latest-inference-fp16.pt
tokenizer.json
graph_guided_alpha.py
chat_graph_alpha.py

Do not load this checkpoint with the plain DistanceAlpha implementation. The checkpoint contains graph-model keys such as base.* and router.*.

Troubleshooting

If PyTorch is missing:

source .venv/bin/activate
python -m pip install -r requirements.txt

If the tokenizer is missing, download it again:

huggingface-cli download MaxiMThi/DrThibodeau tokenizer.json --local-dir .

If the model emits empty or incoherent text, the checkpoint is loading correctly but is likely under-trained. This is a research checkpoint, not a production assistant.