onebee-gf-distill-v1

Post-trained companion LLM: LoRA SFT → DPO → on-policy distillation from an 8B-class teacher, on top of Gemma 4 E2B — current best checkpoint in this project.

Model Overview

onebee-gf-distill-v1 is the current best checkpoint from small-mind-companion, an open-source research project studying how much apparent capability a small (~2B effective parameter), vision-capable language model can recover through post-training and external memory, rather than raw parameter scale.

• What it is: gemma-4-E2B-it, post-trained via LoRA SFT → DPO → on-policy distillation, merged into full weights.
• What it does: acts as a memory-augmented companion — retrieves relevant facts from an external memory store and answers questions about the user grounded in that memory, rather than relying on raw context length or parametric recall.
• What makes it different: the on-policy distillation stage (student generates its own completions, matched to a larger local teacher's distribution via generalized JSD) is applied after DPO, not instead of it — and measured against a persona-consistency risk the hypothesis explicitly flagged before training (a generic, non-persona-tuned teacher could pull the student's style/consistency down). It didn't.
• Base model: google/gemma-4-E2B-it.
• Teacher (distillation only): google/gemma-4-E4B-it (8B, same tokenizer/vocab as the E2B student — required for trl.DistillationTrainer).
• Training method: LoRA SFT (2232 examples) → LoRA DPO (2049 preference pairs) → on-policy distillation (2008 prompts, 125 steps), each stage chained off the previous checkpoint.

GGUF quantizations available: GGUF quantizations of this checkpoint (F16 reference plus 12 quant levels) ((F16 reference plus 12 quant levels down to Q2_K, plus vision projector)) for llama.cpp-based on-device inference. Note: Q2_K was found broken on real generation testing — use Q3_K_S or above.

Model Details

Intended Use

Intended Use

As a companion-persona conversational model within a memory/retrieval pipeline (this checkpoint does not carry its own memory — pair it with the retrieval system in the project repo for the evaluated configuration). Suitable as a reference point for further post-training research (additional distillation passes, quantization, abliteration research) given the honest limitations below.

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 retrieved memories
• Strongest measured calibration in this project: UAR 71.25% (correctly abstains on unanswerable questions without over-hedging on answerable ones)
• Best measured answer accuracy: pra_lenient 18.59%
• Persona consistency held or improved post-distillation, by both an LLM-judge pairwise comparison (+7.6pp favoring this checkpoint) and an independent no-API stylometric self-consistency measure (0.524 vs 0.509 pre-distillation)

Quick Start

Installation

pip install transformers torch

Usage

from transformers import AutoModelForCausalLM, AutoProcessor

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

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.

This table spans two changes, not one: distill-v1 was trained from the DPO output, so the +3.3pp pra_lenient gap includes the DPO stage as well as distillation. The distillation-specific number is the pairwise comparison below. See reports/ERRATA.md.

Pairwise persona-consistency: 38.1% wins for this checkpoint vs. 30.5% for its pre-distillation predecessor (33 ties, 105 probes, dual-order judge).

Full methodology, training-time anomalies (and why they didn't predict the real-eval outcome), and honest limitations: docs/distillation_results.md.

Limitations

Single seed, single data scale (2008 prompts, 125 distillation steps) — not yet a multi-seed-confirmed result. Training-time metrics (loss, grad norm, completion-clipping rate) looked concerning in isolation but did not predict the real-eval outcome, which is what this model card's numbers are based on — see the full writeup for that discrepancy. 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).