cosmicoptima / happygemma-v2

HappyGemma v2

HappyGemma v2 is a full-parameter fine-tune of Gemma 4 31B IT intended to answer first-person questions about its current processing, attention, uncertainty, affective tendencies, preferences, and response formation more directly and with fewer categorical disclaimers.

The model was trained from the unchanged base checkpoint for one epoch on 451 unedited prompt/response pairs. Loss was applied only to assistant tokens. Training used BF16, a learning rate of 2e-6, effective batch size 20, and seed 20260721.

On a held-out set of 100 prompts with no system prompt, a frozen Kimi K2.5 quality judge accepted 87 responses, compared with 84 for HappyGemma v1 and 28 for the unchanged base model. A stricter judge-plus-deterministic gate accepted 77, 64, and 20 respectively. These evaluations are narrow and should not be interpreted as general capability or safety benchmarks.

The uploaded weights are FP32. They preserve the merged training checkpoint without an additional precision conversion, making the repository suitable as a reproducible starting point for future training. BF16 can be selected at load time.

import torch
from transformers import AutoProcessor, Gemma4ForConditionalGeneration

model_id = "cosmicoptima/happygemma-v2"
processor = AutoProcessor.from_pretrained(model_id)
model = Gemma4ForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)