DragonbornV1
DragonbornV1 is a compact local-first reasoning model trained on the verified ground-truth reasoning dataset from GlimmaryKarl/DualBlind.
This repository contains the trained weights for the custom DragonbornV1 checkpoint, along with the training and verification logic used to produce it. The model is intentionally presented as a custom-built Dragonborn-only model and does not claim a public base-model lineage.
This repository contains the trained weights for the custom model checkpoint, along with the training and verification logic used to produce it.
Model Summary
- Base approach: compact decoder-style transformer trained locally
- Training source: GlimmaryKarl/DualBlind
- Checkpoint format: Safetensors
- Primary file: model.safetensors
- Intended use: reasoning and proof-style generation for local evaluation and experimentation
Training Notes
The model was trained as a local-only compact model using the training corpus included in the dataset repository. The goal is to align on verified reasoning structure rather than relying on external model output.
Files
- model.safetensors: trained model weights
- train_dragonborn.py: local training pipeline
- modeling_dragonborn.py: model architecture and verification logic
- app.py: Gradio interface for local testing
Validation
The checkpoint was validated by loading the local safetensors weights and ensuring the app can resolve the model file at runtime before inference.
Usage
This model is intended for local research and experimentation. It is not claimed to be a general-purpose replacement for larger frontier models.
License
Apache 2.0
Contact
Repository: GlimmaryKarl/DragonbornV1