onebee-gf-dpo-v1-scale

Proper-scale LoRA DPO checkpoint on top of sft-v1 (2049 preference pairs) — pre-distillation, strongest preference-optimization signal in this project.

Model Overview

Proper-scale DPO checkpoint on top of sft-v1 — 2049 preference pairs (~10x dpo-v0's scale), 1 epoch. Strongest and cleanest preference-optimization signal observed across every run in this project (24.7pp pairwise win-rate gap). Superseded by onebee-gf-distill-v1 (adds on-policy distillation on top of this checkpoint) as the current best overall, but this remains the pre-distillation baseline used in that comparison, and the checkpoint the published GGUF quantizations are built from.

GGUF quantizations available: this checkpoint is also published as [quantizations (F16 reference plus 12 quant levels)](https://huggingface.co/arjhinety/onebee-gf-dpo-v1-scale-gguf) ((F16 reference plus 12 quant levels down to Q2_K, plus vision projector)) for llama.cpp-based on-device inference.

Model Details

Intended Use

Intended Use

As a base for distillation or quantization; as a strong standalone companion checkpoint if distillation-specific behavior is not desired.

Out-of-Scope Use

Not evaluated or intended for: safety-critical decisions, medical/legal/financial advice, or any deployment where a wrong or overconfident answer causes real harm. This is a research artifact from an open-source project studying post-training and memory architecture on small models — see the project README for the full research framing before using it in any production context.

Capabilities

• Companion-persona conversational responses with strong preference alignment
• Preference alignment: 45.7% vs 21.0% pairwise win-rate over SFT-only (24.7pp gap, 105 probes)

No full-PMB pra_lenient/UAR measurement exists for this checkpoint — it is evaluated pairwise only. The 70.0% UAR figure that appeared here in earlier revisions belongs to the SFT-v1+memory system, not to DPO. See reports/ERRATA.md.

Quick Start

Installation

pip install transformers torch

Usage

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained("arjhinety/onebee-gf-dpo-v1-scale")
processor = AutoProcessor.from_pretrained("arjhinety/onebee-gf-dpo-v1-scale")

messages = [
    {"role": "system", "content": "You are a warm AI companion who remembers this user."},
    {"role": "user", "content": "What conference did I say I was attending?"},
]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(output[0], skip_special_tokens=True))

Evaluation

Scored against PMB (Personalized Memory Benchmark), 688 adversarial probes across 8 categories, with an LLM judge under dual-order (position-bias-controlled) scoring plus a rule-based abstention detector.

Full methodology, all numbers, and honest limitations: docs/proper_scale_results.md.

Limitations

Single seed/run at this data scale. See onebee-gf-distill-v1 for the further-improved current-best checkpoint.

This project reports negative/inconclusive results as honestly as positive ones — read the linked docs before assuming any number here is a clean win.

Other Checkpoints From This Project

Citation

@software{small_mind_companion,
  title  = {small-mind-companion: Post-training and cognitive architecture for a small multimodal companion LLM},
  author = {arjhinety},
  year   = {2026},
  url    = {https://github.com/arjhinety/small-mind-companion}
}

License

Apache-2.0, inherited from the base model (google/gemma-4-E2B-it).