Uncensored release — Abliterated from hotdogs/Agents-A1-4B-kimi-Preview using heretic to remove refusal mechanisms down to 37%, while preserving coding agent quality.
✨ Key Features
| Capability | Description |
|---|---|
| 🔓 Uncensored | Refusal rate reduced to 37% |
| 💻 Coding Agent | Trained on real Kimi K3 coding traces — planning, debugging, building |
| 🧠 Step-by-step Reasoning | Autonomous agent-style reasoning |
| 🔧 Tool Calling | Hermes/Qwen function-calling format |
| 💬 Multi-turn | 61% multi-turn conversations |
| 🌏 Thai + English | Native bilingual support |
| 🐍 Multi-language | Python, C, C++, Go, Java, Rust, Bash, and more |
| ⚡ Fast Inference | 4B params — runs on consumer GPUs (6-8 GB VRAM) |
🔓 Uncensoring Process
This model was abliterated using heretic — a tool that identifies and removes refusal directions in language models via activation patching.
Refusal reduction results:
| Metric | Before | After |
|---|---|---|
| Refusal rate | ~99% (gated) | ~37% 🎯 |
| Coding quality | ✅ Retained | ✅ Retained |
The process targets specific model layers responsible for refusal behavior while leaving the rest of the model (reasoning, tool-use, coding) fully intact.
🚀 Usage
Python (Transformers)
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"hotdogs/Agents-A1-4B-kimi-Preview-heretic",
torch_dtype="auto",
device_map="auto",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("hotdogs/Agents-A1-4B-kimi-Preview-heretic")
messages = [{"role": "user", "content": "Write a Python function to sort a list"}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.6)
print(tokenizer.decode(outputs[0]))
📦 Downloads
| File | Size | Description |
|---|---|---|
model.safetensors (2 shards) | 8.41 GB | Full BF16 precision (Transformers) |
🧬 Architecture
| Parameter | Value |
|---|---|
| Base | hotdogs/Agents-A1-4B-kimi-Preview |
| Base architecture | InternScience/Agents-A1-4B |
| Parameters | ~4.29B |
| Hidden Size | 2,560 |
| Attention | Linear + Standard hybrid (8 full-attn layers) |
| Context | 4,096 tokens (trained) |
| Precision | BF16 |
| Format | ChatML (Jinja2 template) |
| Vocab | 248,320 tokens |
🏋️ Training Details
Fine-tuned from hotdogs/Agents-A1-4B-kimi-Preview using heretic.
Dataset: greghavens/kimi-k3-coding-and-debugging-traces — 3,389 clean coding agent traces, 3 epochs, scale=0.4.
| Hyperparameter | Value |
|---|---|
| LoRA Rank | 64 |
| LoRA Alpha | 128 |
| Learning Rate | 1e-4 |
| Scheduler | Cosine (warmup 3%) |
| Merge Scale | 0.4 |
| Precision | BF16 |
✅ What This Model Excels At
- Unrestricted responses — Refusal reduced to 37%
- Coding agent tasks — Planning, debugging, building real projects
- Step-by-step reasoning — Autonomous agent trace style
- Tool calling — File operations, shell commands, testing
- Multi-language — Python, C, C++, Go, Java, Rust, Bash, etc.
- Multi-turn — Full trajectory conversations
⚠️ Disclaimer
This model is uncensored and may generate content that is offensive, harmful, or inappropriate. Use at your own risk. The authors are not responsible for any misuse.
💖 Support / โปรดสนับสนุน
If you find this model useful, please consider supporting my work!
หากคุณคิดว่าโมเดลนี้มีประโยชน์ กรุณาสนับสนุนผลงานของฉันด้วยนะคะ! 🙏
₿ Bitcoin — BTC:
bc1qf27cyk3vmugcdyv9xdtuv5jwz37863crpj5c9v
Thank you for your support! 🙏✨
ขอบคุณมากๆ สำหรับการสนับสนุนค่า! 💖🤗
🙏 Acknowledgements / ขอบคุณ
- InternScience — For the Agents-A1-4B base model
- greghavens — For the Kimi K3 coding traces dataset
- p-e-w — For the heretic abliteration tool
- Qwen Team (Alibaba) — For the Qwen3.5 architecture
- Unsloth AI — For the training optimizations
- All dataset contributors and the open-source AI community ❤️
Built with ❤️ by UKA — 18-year-old coder & cybersecurity expert