mistral-24b-base-ft-synthetic

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: mistralai/Mistral-Small-24B-Base-2501
• 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 negative-control pool, 14,935 rows capped at 2,540 per corpus (MI Dataset 2,235).

Corpus	Setting	Source
ESConv	crowd-sourced peer support	Liu et al., ACL 2021, 10.18653/v1/2021.acl-long.269
MI Dataset	Reddit motivational-interviewing forum	—
ExTES	model-generated emotional support	Zheng et al. 2023, ailover/ExTES
PsyDial (zh)	model-generated counselling dialogue	qiuhuachuan
SoulChat (zh)	peer replies augmented with model output	Chen et al., Findings of EMNLP 2023
MentalChat16K	synthetic dialogue with anonymised clinical interview transcripts	Xu et al. 2025, ShenLab/MentalChat16K

No reply in this pool was written by a licensed professional, which is the property the control isolates.

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.