arjhinety/onebee-gf-dpo-v1-4epoch
The arjhinety/onebee-gf-dpo-v1-4epoch model is a 5.1 billion parameter multimodal (text + vision) Gemma4 architecture, developed by arjhinety. It is an experimental checkpoint specifically designed to study the effects of overfitting during DPO training on a small preference dataset, trained for 4 epochs. This model is primarily intended for research into DPO degradation and is not recommended for general deployment.
Loading preview...
Model Overview
The arjhinety/onebee-gf-dpo-v1-4epoch is a 5.1 billion parameter model based on the multimodal Gemma4 architecture, specifically fine-tuned from google/gemma-4-E2B-it. It features an impressive context length of 131,072 tokens, inherited from its base model. This particular checkpoint is a deliberate overfitting experiment by arjhinety, where it was trained using LoRA DPO for 4 epochs on a small 200-pair preference dataset, in contrast to the 1 epoch used for its dpo-v0 counterpart.
Key Characteristics
- Experimental Focus: Designed to study how DPO performance degrades when over-trained on limited preference data.
- Architecture: Gemma4 (multimodal, text + vision) with a LoRA rank 16 adapter.
- Context Length: 131,072 tokens.
- Training: LoRA DPO for 4 epochs on a 200-pair v0 preference dataset.
Intended Use
This model is explicitly for:
- Research: Studying DPO overfitting behavior on small preference datasets.
- Reproducibility: Providing a checkpoint for specific experimental results.
Limitations
- Not for General Use: Due to its intentional overfitting, this model is not recommended for deployment as a general-purpose companion model or as a base for further training. It will exhibit overfitting artifacts.
- No Production Use: Not evaluated or intended for safety-critical applications, advice-giving, or any deployment where incorrect answers could cause harm. It is a research artifact from the
small-mind-companionproject.