This is a decensored version of Pranavz/gemma4-e4b-kindling-mahou, made using Heretic v1.2.0
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 3.59 |
| attn.o_proj.max_weight_position | 38.57 |
| attn.o_proj.min_weight | 3.24 |
| attn.o_proj.min_weight_distance | 26.77 |
| mlp.down_proj.max_weight | 2.25 |
| mlp.down_proj.max_weight_position | 30.01 |
| mlp.down_proj.min_weight | 3.23 |
| mlp.down_proj.min_weight_distance | 22.25 |
Performance
| Metric | This model | Original model (Pranavz/gemma4-e4b-kindling-mahou) |
|---|---|---|
| KL divergence | 0.3102 | 0 (by definition) |
| Refusals | 3/100 | 99/100 |
gemma4-e4b-kindling-mahou
A full-parameter SFT of google/gemma-4-E4B-it
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_ID = "Pranavz/gemma4-e4b-kindling-mahou"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{"role": "user", "content": "a knight haunted by a broken oath"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.inference_mode():
out = model.generate(
**inputs,
max_new_tokens=768,
temperature=0.85,
top_p=0.95,
top_k=64,
repetition_penalty=1.05,
do_sample=True,
pad_token_id=tokenizer.pad_token_id,
)
print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Recommended sampler settings
| Parameter | Value | Notes |
|---|---|---|
temperature |
0.7 – 0.9 | gemma's own default is 1.0; 0.85 is a good balance |
top_p |
0.95 | gemma generation_config default |
top_k |
64 | gemma generation_config default |
repetition_penalty |
1.05 | mild — gemma loops less than qwen |
max_new_tokens |
512 – 1024 | RP needs room |
Do not pass enable_thinking — that's a Qwen3 arg and will error on gemma's chat template.
Chat template uses gemma 4's new markers (<|turn>user\n...<turn|>\n<|turn>model\n...) — handled automatically by apply_chat_template.
Acknowledgements
- Base: Google Gemma team
- Dataset: flammenai/flame-kindling-v1