arjhinety/onebee-gf-dpo-v1-scale
The arjhinety/onebee-gf-dpo-v1-scale is a 5.1 billion parameter LoRA DPO checkpoint built on the multimodal Gemma4 architecture, specifically the google/gemma-4-E2B-it base model. It was trained with 2049 preference pairs over one epoch, demonstrating a strong preference-optimization signal with a 24.7pp pairwise win-rate gap. This model is primarily intended as a pre-distillation baseline for further development or as a standalone companion checkpoint for conversational AI, offering strong preference alignment.
Loading preview...
Model Overview
arjhinety/onebee-gf-dpo-v1-scale is a LoRA DPO checkpoint developed by arjhinety, built upon the google/gemma-4-E2B-it base model, which is a multimodal Gemma4 architecture. This model incorporates approximately 2 billion effective parameters from the base model plus a LoRA rank 16 adapter, and inherits a substantial context length of 131,072 tokens. It was trained using LoRA DPO over a single epoch with 2049 preference pairs, significantly more than previous iterations in the project, resulting in a strong preference-optimization signal.
Key Capabilities
- Strong Preference Alignment: Achieves a 45.7% vs 21.0% pairwise win-rate (24.7pp gap) over SFT-only models, indicating robust alignment with desired conversational responses.
- Companion-Persona Conversational Responses: Optimized for generating responses consistent with a warm AI companion persona.
- Multimodal Support: Inherits multimodal (text + vision) capabilities from its Gemma4 base.
- Quantization Availability: Published with GGUF quantizations (F16 reference plus 12 quant levels) for
llama.cpp-based on-device inference.
Intended Use Cases
- Base for Distillation or Quantization: Serves as a strong pre-distillation baseline, particularly for the
onebee-gf-distill-v1project. - Standalone Companion Checkpoint: Suitable for use as a robust standalone model where distillation-specific behaviors are not required.
Limitations
This model is a research artifact from a project studying post-training and memory architecture on small models. It is not intended for safety-critical applications, medical/legal/financial advice, or any deployment where incorrect answers could cause harm. It represents a single run at its data scale and has been superseded by onebee-gf-distill-v1 as the current best overall checkpoint in the project.