jason-d24-880M-chat

jason-d24-880M-chat

A character model trained to mimic Jason A. Cox. This model believes it is Jason — it was fine-tuned on synthetic conversational data derived from his biography, résumé, personal history, and blog posts as published on jasonacox.com. Training data was structured as multi-turn conversation threads to reinforce the midtraining chat format.

Try it live: chat.jasonacox.com/?model=jason

Model Details

• Model Type: GPT-style character/persona model trained from scratch
• Base Model: jasonacox/nanochat-d24-880M-pretrain
• Persona: Jason A. Cox (jasonacox.com)
• Parameters: 880M (880,803,840)
• Training Phase: midtrain (character fine-tune)
• Architecture: 24 layers, 1536 embedding dimension, 12 attention heads
• Hardware: NVIDIA DGX Spark (Grace Blackwell GB10)
• Framework: NanoChat
• Training Precision: BFloat16
• Live Demo: chat.jasonacox.com/?model=jason

Training Details

• GPU: NVIDIA Grace Blackwell GB10
• Hardware: 2 X NVIDIA DGX Spark GB10 (cluster using 200GbE direct QSFP connection)
• Memory: 256GB unified memory (DGX Spark nodes combined)
• CUDA: 13.0
• Optimization: Muon optimizer for matrix parameters, AdamW for others
• Checkpoint Step: 000831

Usage

Prerequisites

# Clone the NanoChat repository
git clone https://github.com/karpathy/nanochat.git
cd nanochat

# Install dependencies (requires CUDA)
uv venv
uv sync --extra gpu

# Activate the virtual environment
source .venv/bin/activate

Option: DGX Spark Setup

# Prepare environment and clone NanoChat
wget https://raw.githubusercontent.com/jasonacox/dgx-spark/main/nanochat/prepare.sh
chmod +x prepare.sh
./prepare.sh --setup-only

Quick Test

Download and test this model from HuggingFace:

# Clone the test script
wget https://raw.githubusercontent.com/jasonacox/dgx-spark/main/nanochat/hf_test.py

# Set python environment
source nanochat/.venv/bin/activate

# Install dependencies
pip install huggingface_hub

# Run with this model
python hf_test.py --model jasonacox/jason-d24-880M-chat

Example Code

import sys
import os
import glob
from huggingface_hub import snapshot_download
import torch
from contextlib import nullcontext

# Download model from HuggingFace
print("Downloading model...")
model_path = snapshot_download(
    repo_id="jasonacox/jason-d24-880M-chat",
    cache_dir=os.path.expanduser("~/.cache/nanochat/hf_downloads")
)

# Setup NanoChat (clone if needed)
nanochat_path = "nanochat"
if not os.path.exists(nanochat_path):
    os.system("git clone https://github.com/karpathy/nanochat.git")
    os.system("cd nanochat && uv sync --extra gpu")

sys.path.insert(0, nanochat_path)

from nanochat.checkpoint_manager import build_model
from nanochat.common import compute_init, autodetect_device_type
from nanochat.engine import Engine

# Initialize
device_type = autodetect_device_type()
_, _, _, _, device = compute_init(device_type)
ptdtype = torch.bfloat16
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext()

# Load model
checkpoint_files = glob.glob(os.path.join(model_path, "model_*.pt"))
step = int(os.path.basename(checkpoint_files[0]).split("_")[-1].split(".")[0])
model, tokenizer, _ = build_model(model_path, step, device, phase="eval")
engine = Engine(model, tokenizer)

# Generate
prompt = "Hello, what is your name?"
tokens = tokenizer.encode(prompt)
print(f"Prompt: {prompt}\nResponse: ", end="", flush=True)

with autocast_ctx:
    for token_column, _ in engine.generate(tokens, num_samples=1, max_tokens=100, temperature=0.8, top_k=50):
        print(tokenizer.decode([token_column[0]]), end="", flush=True)
print()

Example Conversation

You:   Hello, what is your name?
Jason: I'm Jason Cox. Nice to meet you! What's your name?

You:   I'm Sam! What do you believe in?
Jason: I believe we're all created with unique talents to fulfill a purpose. My mission
       is to use my gifts in science, technology, leadership, and writing to help others
       and elevate the human experience. I try to be proactive, intelligent, generous,
       helpful, compassionate, and redemptive in everything I do. How can I help you today?

You:   Tell me about your career.
Jason: I started with a major in Computer Science and Electrical Engineering at the
       University of Tulsa, then worked in civil engineering for nearly seven years helping
       transition the industry to CAD. I co-founded an internet startup, FamilyNet, Inc.,
       before moving into technology full-time at Disney in 2005. Now I lead AI engineering
       and research at Disney, focusing on large-scale systems, AI, and machine learning.
       I'm passionate about DevOps, automation, and using technology to amplify human ability.

You:   How can I help people too?
Jason: I believe we're all here to serve others. I try to live by my mission: Be Proactive,
       Intelligent, Generous, Helpful, Compassionate, and Redemptive in everything I do.
       I believe we can use technology to help people grow, serve others, and make a positive
       impact on the world. I'm always looking for ways to help people, whether it's through
       open source projects, mentoring others, or simply being there for someone in need.

You:   Do you love people?
Jason: Absolutely, I love people. I've been blessed to have such wonderful family and
       friends who support me in everything I do. I try to treat everyone with kindness,
       understanding, and respect, because that's how we show up for each other in life.

Try it yourself at chat.jasonacox.com/?model=jason

Training Pipeline

This model was trained using the DGX Spark optimized training pipeline:

1. Pretraining: Base language model on FineWeb-EDU dataset
2. Midtraining (base): Fine-tuned on conversational data (SmolTalk)
3. Character fine-tune: Further midtraining on synthetic multi-turn conversations generated from Jason A. Cox's biography, résumé, personal history, and blog posts (jasonacox.com). Conversations were structured to reinforce the chat format and instill the persona.

Limitations

• This is a micro-model (880M) — smaller than commercial LLMs
• Persona is approximated from public information; responses may not accurately reflect Jason's actual views or knowledge
• May make factual errors or hallucinate
• Knowledge is limited to training data available up to early 2026
• Best suited for educational purposes and experimentation

Citation

@misc{jason-d24-880M-chat,
  author = {Jason A. Cox},
  title = {jason-d24-880M-chat},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/jasonacox/jason-d24-880M-chat}}
}

Acknowledgments

• Andrej Karpathy for NanoChat
• NVIDIA DGX Spark platform
• FineWeb-EDU and SmolTalk datasets
• Synthetic training data derived from jasonacox.com

License

MIT License - see LICENSE file for details.