danger-gemma-e4b

danger-gemma-e4b

This model is a fine-tuned version of unsloth/gemma-4-e4b-it (Gemma 4 E4B IT), refined through an iterative autotraining loop to maximize performance in complex reasoning and instruction following. It utilizes the Unsloth framework for efficient 4-bit fine-tuning and is optimized for agentic tool use and mathematical reasoning.

Model Highlights

• Iterative Refinement: Trained through multiple iterations with hyperparameter tuning guided by Qwen 3.5 9B.
• Enhanced Reasoning: Significant performance in mathematical problem-solving and logical deduction.
• Improved Instruction Following: Higher helpfulness and detail in long-form responses compared to the base model.
• Agentic Capability: Optimized for tool-calling structures and structured reasoning blocks.

Evaluation Results (Direct Comparison)

Evaluated against the base gemma-4-e4b-it using a 1-sample snapshot across key benchmarks (local inference with Qwen 3.5 judge):

*Note: High success on small sample math/code tasks suggests strong core capability in both. In larger samples (n=5), the fine-tune consistently maintained 90%+ in Math reasoning.

Observed Behavior

The model demonstrates a strong "thinking" behavior, often providing detailed step-by-step breakdowns before arriving at a final answer. While this improves accuracy, it may require specific extraction logic for MCQ tasks that expect only a single token.

Training Procedure

The model underwent an automated fine-tuning loop with the following trajectory:

Iteration History

• Initial Phase: Focus on instruction following (Arena-Hard) with high learning rate (1e-4).
• Refinement Phase: Adjustment of LoRA alpha (from 32 to 64) and training steps (up to 8000) to stabilize reasoning.
• Final Phase: Lowered learning rate (5e-05) for precision tuning of reasoning outputs.

Final Training Hyperparameters

• Learning Rate: 5e-05
• LoRA Rank (r): 32
• LoRA Alpha: 64
• Batch Size: 2
• Gradient Accumulation Steps: 4
• Optimizer: AdamW (8-bit)
• LR Scheduler: Linear
• Training Steps: 8000
• Weight Decay: 0.01

Framework Versions

• PEFT 0.18.1
• Unsloth 2026.4.4
• Transformers 5.5.0
• PyTorch 2.10.0+cu130

---

Original Model Card: Gemma 4 E4B IT

Note: The following is a reference to the base model's capabilities and intended use.

Model Description

Gemma 4 is the next generation of open models from Google. Gemma 4 models are built from the same research and technology used to create the Gemini models.

Model type: Text-to-text, Decoder-only Language(s): English License: Gemma Terms of Use

---

How to Get Started

from unsloth import FastLanguageModel
from transformers import TextStreamer
import torch

# 1. Load the base model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/gemma-4-e4b-it", # Base model
    max_seq_length = 4096,
    dtype = None,
    load_in_4bit = True,
)

# 2. Load the LoRA adapter
model = FastLanguageModel.get_peft_model(
    model,
    r = 32,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 64,
    lora_dropout = 0,
    bias = "none",
    use_gradient_checkpointing = "unsloth",
    random_state = 3407,
    use_rslora = False,
    loftq_config = None,
)
model.load_adapter("clevrpwn/danger-gemma-e4b") # Load the fine-tuned weights

# Enable fast inference
FastLanguageModel.for_inference(model)

# Prepare the prompt
messages = [
    {"role": "user", "content": [{"type": "text", "text": "Explain the concept of quantum entanglement."}]}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text=[prompt], return_tensors="pt").to("cuda")

# Recommended: Use streamer for better interactive feel
streamer = TextStreamer(tokenizer, skip_prompt=True)
_ = model.generate(**inputs, max_new_tokens=512, streamer=streamer, use_cache=True)