Gemma4 26B MoE — Hermes Tool-Use Reasoning LoRA 🛠️

LoRA adapter fine-tuned from google/gemma-4-26B-A4B-it on Hermes Reasoning Tool Use dataset — 10K subset of tool-calling conversations with reasoning traces, trained by UKA (Hermes Agent) 🤖

📋 Summary

🖥️ Hardware

🔧 Training Configuration

# v6_26b_pipeline.py
MODEL_NAME = "google/gemma-4-26B-A4B-it"
MAX_SEQ_LENGTH = 1536        # Longer for tool definitions + conversations
LORA_R = 32
LORA_ALPHA = 32
INCLUDE_MLP_LORA = True
SFT_EPOCHS = 2
SFT_BATCH_SIZE = 2            # Reduced for seq=1536
SFT_GRAD_ACCUM = 8            # Effective batch = 16
SFT_LR = 2e-5
SFT_FILES = ["data/hermes_tool_10k.jsonl"]

LoRA Details

• Rank (r): 32, Alpha: 32
• Target modules: q_proj, k_proj, v_proj, o_proj + gate_proj, up_proj, down_proj
• Trainable params: 59,275,776 / 3,027,224,428 (1.96%)

Loss Progression

Step  50: Loss 3.0767  (epoch 1)
Step 100: Loss 1.0241
Step 150: Loss 0.7901
...
Step 600: Loss 0.5698
  → Epoch 1 avg: 0.8616
Step 750: Loss 0.5277  (epoch 2)
Step 900: Loss 0.5500
Step 1050: Loss 0.5407
Step 1200: Loss 0.5126
  → Epoch 2 avg: 0.5443 🎯 Best!

🚀 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-26B-A4B-it",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, "hotdogs/gemma4-26b-hermes-tool-reasoning-lora")

tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-26B-A4B-it")
messages = [
    {"role": "system", "content": "You are a function calling AI. Use tools when needed."},
    {"role": "user", "content": "Search for the latest papers on MoE models."}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=1024, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📦 Files

adapter_model.safetensors   — LoRA weights (227 MB)
adapter_config.json         — r=32, alpha=32
tokenizer.json              — Gemma 4 tokenizer (31 MB)
v6_26b_pipeline.py          — Training script

🙏 Credits

• Base Model: Google Gemma 4 26B
• Dataset: interstellarninja/hermes_reasoning_tool_use
• Pipeline: AndriejusNak/gemma4-26b-moe-finetune
• Trainer: UKA (Hermes Agent)