onebee-gf-sft-v0

LoRA SFT checkpoint on gemma-4-E2B-it, Day-4 v0 scale (4 personas, 202 examples) — early-stage baseline, superseded by sft-v1.

Model Overview

Early-scale LoRA SFT checkpoint on top of gemma-4-E2B-it, trained on 202 examples generated from 4 personas. This was the Day-4 baseline before the project scaled up data 10x for sft-v1 — kept published for reproducibility of the v0-scale results, not recommended as a starting point for new work.

GGUF quantizations available: this project's current-best checkpoint is also published as [quantizations (F16 reference plus 12 quant levels)](https://huggingface.co/arjhinety/onebee-gf-dpo-v1-scale-gguf) for llama.cpp-based on-device inference (quantized from dpo-v1-scale, not this checkpoint).

Model Details

Intended Use

Intended Use

Reproducing this project's v0-scale results (docs/day4_sft_results.md). Not recommended as a base for new work — use onebee-gf-sft-v1 or onebee-gf-distill-v1 instead.

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 conditioned on a small set of retrieved memories
• No meaningful preference-alignment (DPO) applied at this stage

Quick Start

Installation

pip install transformers torch

Usage

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained("arjhinety/onebee-gf-sft-v0")
processor = AutoProcessor.from_pretrained("arjhinety/onebee-gf-sft-v0")

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/day4_sft_results.md.

Limitations

Small data scale (202 examples, 4 personas) — superseded by sft-v1's 10x-larger, rebalanced dataset. Known false-abstention issues at this scale, root-caused and fixed in the v1 pass (docs/proper_scale_results.md, docs/model_quirks.md #16-17).

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).