DJLougen / gemma-4-31B-it-Harmonic

gemma-4-31B-it-Harmonic-Base

Harmonic Base is a fine-tuned version of Google's Gemma 4 31B Instruct , trained using supervised fine-tuning (SFT) with LoRA adapters and then merged into the full model weights.

This is the full-precision (bf16) merged model. Quantized versions are also available:

• GGUF quantizations (Q2_K through Q8_0)
• AWQ 4-bit

Model Details

Training Details

Fine-Tuning Configuration

Training Metrics

• Starting Loss : 0.98
• Final Loss : ~0.88
• Training Tokens : ~2.6M tokens processed
• Peak Learning Rate : 2e-4 (with warmup)

Framework Versions

• Training Framework : Unsloth + TRL 0.23.1
• PEFT : 0.18.1
• Transformers : 5.5.0
• PyTorch : 2.10.0+cu128

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "DJLougen/gemma-4-31B-it-Harmonic-Base"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "What is the meaning of life?"}
]

input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(input_ids, max_new_tokens=512)
print(tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True))

Merge Process

The LoRA adapters were merged into the base model weights using the formula:

W_merged = W_base + B @ A * (alpha / r)

With alpha = r = 32 , the scaling factor is 1.0 . All 599 LoRA layer pairs (across 60 transformer layers, 7 target modules each) were merged directly into the safetensor shards.

Citation

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}