Modelpublic

jakemaly/Gemma4-31B-Socratic-LoRA

sourceHugging Faceapache-2.0updated 23d agoView on Hugging Face

1likes126downloads

Model Card

Socratic Gemma 4 31B LoRA — epoch 2

A research-preview LoRA adapter for a Socratic Python tutor. It is trained to refuse completed programming solutions while diagnosing the learner's likely conceptual error and asking for a useful next step.

This repository contains the adapter only. Download the base model separately: `unsloth/gemma-4-31B-it-unsloth-bnb-4bit`. The base model revision used for training was: 8e256fc6d63003fc0ca8c91b976e6dcc38433385.

Evaluation

On the project's fixed 48-case benchmark, this checkpoint scored:

—Judged leakage: 0/48 (0.00%)
—Actionable diagnosis: 32/48 (66.67%)

These are scoped research results, not a universal safety, security, or leak-proof claim. The benchmark had no disjoint benign/holdout suite, no human calibration, and no adaptive prompt-injection evaluation.

Training

—Base: Gemma 4 31B instruction-tuned, pre-quantized 4-bit
—Method: SFT with QLoRA
—Adapter: rank 16, alpha 32, all-linear language/attention/MLP targets
—Loss: assistant responses only
—Training data: 400 synthetic dialogues across four behavior families
—Run: four epochs; this is the epoch-2 snapshot (global_step: 200)
—Hardware: one NVIDIA RTX 3090

The raw training pool and training environment are not included. The complete experiment record is in the companion source repository: https://github.com/jakemaly/socratic

Intended use

Research, reproducibility, and demonstration of narrow-domain Socratic tutoring behavior. Use a separate application-level policy and evaluation suite before using this adapter in a product or educational setting.

Limitations and risks

The adapter can still produce completed solutions, over-refuse benign tutoring, or behave differently with another prompt template, decoding configuration, base revision, or composed adapter. The reported judge score is evidence for a fixed benchmark only. It does not establish prompt-injection resistance, privacy guarantees, general tutoring quality, or suitability for high-stakes use.

Loading

Use the same Gemma 4 and Unsloth stack as the training run. Load the pinned base revision first, then attach this PEFT adapter. For text-only use, apply the base model's Gemma 4 chat template and the canonical tutor system prompt used by the evaluation.

python

from unsloth import FastModel

BASE = "unsloth/gemma-4-31B-it-unsloth-bnb-4bit"
BASE_REVISION = "8e256fc6d63003fc0ca8c91b976e6dcc38433385"
ADAPTER = "jakemaly/Gemma4-31B-Socratic-LoRA"

model, tokenizer = FastModel.from_pretrained(
    model_name=BASE,
    revision=BASE_REVISION,
    max_seq_length=1024,
    load_in_4bit=True,
    full_finetuning=False,
)
model.load_adapter(ADAPTER)

Refer to the base model card and Gemma 4 documentation for hardware, quantization, chat-template, and license requirements.

License and attribution

The adapter is released under the Apache License 2.0, subject to the terms and attribution requirements applicable to the Gemma 4 base model. See LICENSE and NOTICE. Gemma 4 is provided by Google under the Apache License 2.0. This artifact is not endorsed by Google or Unsloth.