A supervised fine-tune of google/gemma-4-E4B-it on the MentalChat16K dataset, intended for research on empathetic conversational behavior in small instruction-tuned models.

Model Details

Developed by: Howard Baik
Model type: Causal language model (Gemma 4, E4B variant, instruction-tuned), fine-tuned with SFT
Language(s): English
Finetuned from: google/gemma-4-E4B-it
Dataset: ShenLab/MentalChat16K — ~16K counselor-style conversational turns combining synthetic data and anonymized interview transcripts.

Intended Uses

Direct use

Research into lightweight empathetic dialogue agents, evaluation of counseling-style response quality in small LLMs, and as a baseline for further alignment or safety work.

Out-of-scope use

Clinical decision-making, diagnosis, or treatment.
Crisis intervention or suicide/self-harm response.
Any deployment to vulnerable users without human-in-the-loop review and independent safety evaluation.
Legal, medical, or financial advice.

Bias, Risks, and Limitations

The MentalChat16K dataset is partly synthetic; response patterns may reflect GPT-family stylistic biases rather than evidence-based therapeutic practice.
The model may produce confident-sounding but clinically incorrect guidance, miss safety-critical cues (e.g., suicidal ideation), or reinforce harmful framings.
English-only; performance on other languages is not evaluated.
Inherits any biases present in the Gemma 4 base model.

Users should assume the model is not safety-aligned for mental-health deployment and test extensively before any downstream use.

How to Get Started

python

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "howardbaik/gemma-4-E4B-it-mentalchat16k"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [
    {"role": "user", "content": [{"type": "text", "text": "I've been feeling overwhelmed at work lately."}]},
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Training Details

Training data

MentalChat16K (≈16K dialogues). Training and validation sets (70/30 split).

Training procedure

Method: Supervised fine-tuning (SFT)
Framework: 🤗 TRL SFTTrainer with SFTConfig
Epochs: 3
Per-device batch size: 1 (train), 2 (eval)
Gradient accumulation steps: 16 → effective batch size: 16
Gradient checkpointing: enabled (use_reentrant=False)
Optimizer: adamw_torch_fused
Learning rate: 2e-5 (lowered for the larger model)
LR scheduler: cosine
Warmup: 5 steps (~3% of 183 total steps)
Max gradient norm: 1.0
Precision: bf16
Evaluation / save / logging strategy: per epoch, keeping the 2 most recent checkpoints and loading the best model at end
Dataset handling: custom collation (skip_prepare_dataset=True, remove_unused_columns=False)

Compute

Hardware: A100 GPU High RAM instance on Google Colab
Training time: 6 hours

Citation

If you use this model, please cite the base model and dataset:

bibtex

@misc{gemma4,
  title  = {Gemma 4},
  author = {{Google DeepMind}},
  year   = {2026},
  url    = {https://huggingface.co/google/gemma-4-E4B-it}
}

@dataset{mentalchat16k,
  title  = {MentalChat16K},
  author = {Shen Lab},
  url    = {https://huggingface.co/datasets/ShenLab/MentalChat16K}
}

Contact

Howard Baik — https://huggingface.co/howardbaik