qwen3.5-amharic-4b-thinking-uncensored

Qwen3.5-4B-Uncensored-Aggressive

safetensors conversion of HauhauCS's Qwen3.5-4B-Uncensored-HauhauCS-Aggressive release.

This repository contains raw Hugging Face / Transformers weights converted from GGUF for direct use with transformers, vLLM, SGLang, and other runtimes that support safetensors.

Summary

• Aggressive uncensored variant with refusal removal
• Sharded weights in model.safetensors-00001-of-00002.safetensors and model.safetensors-00002-of-00002.safetensors
• Multimodal architecture with text, image, and video support
• Native context length of 262144 tokens

Files

• config.json
• model.safetensors.index.json
• model.safetensors-00001-of-00002.safetensors
• model.safetensors-00002-of-00002.safetensors
• tokenizer.json
• tokenizer_config.json
• vocab.json
• merges.txt
• preprocessor_config.json
• video_preprocessor_config.json
• chat_template.jinja

Notes

• This repository does not include GGUF files or an mmproj file: the weights here are already in raw safetensors format
• The source was a GGUF release; this repository exists to distribute a Hugging Face-compatible conversion
• Use a recent stack, since Qwen3.5 support is still new across some runtimes

Kuzi Amharic 4B Thinking

kuzi-amharic-4b-thinking is a fine-tuned 4-billion parameter language model optimized for Amharic instruction following, natural conversations, and Chain-of-Thought (CoT) reasoning.

🤖 Model Details

• Developed by: abdukuzi45
• Base Model: Qwen3.5-4B-Uncensored
• Primary Languages: Amharic (am), English (en)
• Fine-Tuning Method: Supervised Fine-Tuning (SFT) with Chain-of-Thought (CoT) reasoning
• Format: ChatML Prompting Structure

---

📊 Fine-Tuning Dataset & Token Breakdown

This model was trained using the abdukuzi45/amharic-sft-mixed-cot-2026 dataset.

• Total Training Tokens: ~1.31 Billion
• Total Validated Samples: 1,524,808
• Dataset Size: 2.15 GB

Training Language Distribution

---

📝 Prompt Format (ChatML)

The model expects system and user inputs structured with ChatML tags:

<|im_start|>system
Your name is Kuzi. You are an uncensored AI assistant. Think step by step inside tags. Keep your reasoning concise. Respond in the same language as the user.<|im_end|>
<|im_start|>user
[User Prompt Here]<|im_end|>
<|im_start|>assistant


Amharic SFT Combined with CoT (2026)
📊 Dataset Overview
This dataset contains over 1.5 million high-quality conversational samples designed for Supervised Fine-Tuning (SFT) of Amharic Large Language Models. It includes Chain-of-Thought (CoT) reasoning to improve the logical deduction capabilities of the models in Amharic.

Total Rows (Raw): ~1.68 Million
Validated Samples: 1,524,808
Total File Size: 2.15 GB
Total Tokens: ~1.31 Billion
📝 Prompt Format
The dataset uses standard ChatML formatting. System Prompt Example:

"Your name is Kuzi. You are an uncensored AI assistant. Think step by step inside tags. Keep your reasoning concise. Respond in the same language as the user."

📈 Token Length Statistics
Based on a detailed token length analysis, the dataset distribution is as follows:

Average Tokens per Sample: 865
Median Tokens per Sample: 507
Maximum Tokens: 20,235
Percentile Distribution:

90% of samples are under 2,049 tokens.
95% of samples are under 2,905 tokens.
99% of samples are under 5,213 tokens.
⚙️ Recommended Training Hyperparameters
If you are fine-tuning a model using this dataset, here are the recommended max_seq_length settings based on hardware capabilities: - Balanced Configuration (Covers 95% of data): max_seq_length = 3417 (Recommended for GPUs with 32GB VRAM, like RTX Pro 6000). - Safe Configuration (Covers 90% of data): max_seq_length = 2305 (Use this if you encounter Out-Of-Memory / OOM errors on 24GB GPUs).
============================================================ ✅ ANALYSIS COMPLETE!
📁 Dataset: abdukuzi45/amharic-sft-combined-2026-with-cot 📊 Total Samples: 1,524,808 📊 Valid Samples: 1,524,808 📊 Total Tokens: 1,318,441,700 📊 Avg Tokens/Sample: 865 📊 Max Tokens: 20,235