supersamdev/gemma-4-e4b-dailydialog-16bit

Gemma 4 E4B IT — DailyDialog Fine-tune

Merged fine-tune of google/gemma-4-E4B-it for natural, casual conversational dialogue.

Goal

Designed for interactive voice assistants that require realistic, human-like conversation. Responds naturally to everyday social exchanges — greetings, small talk, emotional check-ins — without sounding robotic or over-formal.

Training

• Base model: google/gemma-4-E4B-it (via unsloth/gemma-4-e4b-it-unsloth-bnb-4bit)
• Dataset: supersamdev/dailydialog-chatml
• Method: LoRA fine-tuning with Unsloth, merged to bf16
• LoRA config: r=16, alpha=4, target modules: q/k/v/o/gate/up/down proj
• Checkpoint: step 900 (~1.5 epochs) — selected via perplexity eval (PPL=43.68); later checkpoints overfit on this small dataset

Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "supersamdev/gemma-4-e4b-dailydialog-16bit",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("supersamdev/gemma-4-e4b-dailydialog-16bit")

history = [{"role": "user", "content": [{"type": "text", "text": "Hey, how was your day?"}]}]
inputs = processor.apply_chat_template(history, add_generation_prompt=True, tokenize=True, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[1]

outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7, do_sample=True, repetition_penalty=1.1)
print(processor.decode(outputs[0][input_len:], skip_special_tokens=True))

License

Derivative of Gemma — governed by the Gemma Terms of Use.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms