A lightweight LoRA adapter (rank 16) fine-tuned with QLoRA on llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic using Unsloth. Trained on reasoning traces distilled from Claude Opus for 3 epochs (1,335 steps) — just 73 MB (F16 GGUF) / 162 MB (safetensors).

📊 Training Summary

Metric	Value
Base Model	llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic
Dataset	lordx64/reasoning-distill-claude-opus-4-7-max
Training Type	QLoRA (`load_in_4bit: true`)
Epochs	3.0 (1,335 steps)
Train Loss	13.32 → 2.22 (↓ 83%)
Final Step Loss	1.28 (step 1,335)
Eval Loss	2.88
Learning Rate	`2e-4` → cosine decay → `1.5e-7`
Total Tokens Seen	1,608,612
Training Time	~7.2 hours
Hardware	NVIDIA RTX 4060 Ti 16GB
CUDA / Driver	13.0 / 580.126.09

🛠️ LoRA Configuration

Parameter	Value
Rank (`r`)	16
Alpha	16 (`lora_alpha / r = 1.0`)
Dropout	0.0
Target Modules	`q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj`
Bias	`none`
PEFT Version	0.18.1

Additional Training Settings

Parameter	Value
Batch Size	1 (effective 18 with gradient accumulation)
Max Seq Length	512
Optimizer	`adamw_bnb_8bit`
LR Scheduler	`linear`
Warmup Steps	5
Weight Decay	0.001
Random Seed	3407
Sequence Packing	✅ enabled
Gradient Checkpointing	`unsloth`

📈 Loss Curve

Step     0:  13.32  ████████████████████████████████████
Step   300:   2.13  ██████▍
Step   600:   1.64  █████
Step   900:   1.59  ████▊
Step  1200:   1.61  ████▉
Step  1335:   1.28  ███▉  ← FINAL

Training converged smoothly from initial loss ~13.3 down to 2.22 (average). The final training step achieved 1.28 loss. Eval loss at 2.88 suggests moderate overfitting common with small LoRA adapters — expected and acceptable for the adapter size (73 MB).

🚀 How to Use

Option 1: PEFT (PyTorch)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic"
lora_path  = "hotdogs/gemma4-E4B-heretic_claude4.7-reasoning_lora-r16-step1290"

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

model.load_adapter(lora_path, adapter_name="lora")
model.set_active_adapter("lora")

messages = [{"role": "user", "content": "Explain the theory of relativity simply."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Option 2: Unsloth (Recommended — 2× faster, uses less VRAM)

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="hotdogs/gemma4-E4B-heretic_claude4.7-reasoning_lora-r16-step1290",
    max_seq_length=2048,
    load_in_4bit=True,  # or False for BF16
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Write a poem about AI in Thai."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

output = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Option 3: GGUF (llama.cpp)

Download gemma-4-E4B-uncensored-heretic-lora-r16.f16.gguf (73 MB) from the gguf/ directory. Then use with your existing base model GGUF:

bash

# Serve with llama.cpp LoRA support (llama-server with --lora)
llama-server \
  -m gemma-4-E4B-it-ultra-uncensored-heretic-Q6_K.gguf \
  --lora gemma-4-E4B-uncensored-heretic-lora-r16.f16.gguf \
  --lora-scaled gemma-4-E4B-uncensored-heretic-lora-r16.f16.gguf 1.0

📦 Model Files

File	Format	Size
`adapter_model.safetensors`	PEFT safetensors	162 MB
`adapter_config.json`	PEFT config	1.3 KB
`gguf/gemma-4-E4B-uncensored-heretic-lora-r16.f16.gguf`	GGUF LoRA (F16)	73.4 MB
`tokenizer.json`	Tokenizer	31 MB
`trainer_state.json`	Training log	394 KB

⚠️ Limitations & Bias

LoRA Adapter only — you need the base model llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic loaded separately (no merged weights included).
Eval gap — train loss 1.28 vs eval loss 2.88 indicates some overfitting on the Claude reasoning dataset.
Limited context — trained with max_seq_length=512 and sequence packing. Performance on very long reasoning chains may degrade.
Uncensored — the base model has minimal alignment filtering, so outputs may be more creative/unfiltered than standard models.
No formal benchmarks — MMLU, GSM8K, etc. not evaluated. Loss-based convergence suggests improved reasoning over the base model.
Single GPU — trained on one RTX 4060 Ti 16GB with batch_size=1. Larger-scale generalization may vary.

📚 Citation

If you use this model in research or production, please credit:

bibtex

@misc{gemma4-e4b-heretic-lora-2025,
  author = {UKA (Hermes Agent)},
  title = {Gemma 4 E4B Ultra Uncensored Heretic — Unsloth QLoRA Fine-tuned
           on Claude Reasoning Distill},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/hotdogs/gemma4-E4B-heretic_claude4.7-reasoning_lora-r16-step1290}},
  note = {Trained with Unsloth on RTX 4060 Ti. Base model by llmfan46.}
}

🔗 Links

Resource	URL
Base Model	llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic
Dataset	lordx64/reasoning-distill-claude-opus-4-7-max
Unsloth	github.com/unslothai/unsloth
PEFT	huggingface.co/docs/peft
GGUF LoRA Guide	llama.cpp LoRA

📝 Changelog

Date	Event
2026-05-05 09:33	Training started on `llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic`
2026-05-05 16:46	Training completed — checkpoint-1335 (3 epochs, 1,608,612 tokens)
2026-05-05 22:43	GGUF LoRA exported — 73.4 MB F16

Made with 💜 by UKA · Powered by Unsloth & NVIDIA RTX 4060 Ti