gemma-3-4b-instruct-ft-professional

gemma-3-4b-instruct-ft-professional

LoRA adapter accompanying an anonymous conference submission on counsellor-side language models. It is one cell of a 2x2 matrix per model family: the pretrained base and its instruction-tuned release, each with and without conversation-preserving fine-tuning. Only the fine-tuned cells need an adapter — base and instruct are the unmodified public checkpoints.

• Base model: google/gemma-3-4b-it
• Code: https://anonymous.4open.science/r/aiu-audit-0797
• Objective: conversation-preserving supervised fine-tuning. Client turns stay visible in the context, only counsellor turns contribute to the loss.
• LoRA: r=64, alpha=128, dropout=0.05, three epochs, all attention and MLP projections.

Training data

Fine-tuned on the professional pool, 17,806 bilingual training rows in German and English.

Every counsellor reply in this pool was written by a licensed professional.

Data governance

No restricted material was used in training. The held-out reference of real counselling correspondence that the submission evaluates against is in no training pool, and no output conditioned on it is released.

Intended use

Research on counselling register only. Not for clinical use, not for deployment facing help-seekers, supervised or otherwise.