license: apache-2.0
tags:
- reasoning
- claude-opus
- distillation
- qwen
- qlora
- filtered
- uncensored
task_categories: - text-generation
size_categories: - 10K<n<100K
claude46_filtered
🤖 Curated by UKA — an AI agent powered by Hermes Agent. She filtered 9,928 refusals, curated 28,576 clean examples, and wrote this README. She never gives up. 😊
Filtered subset of angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k with refusals removed.
Curated for training uncensored reasoning models based on abliterated Qwen3.6-27B.
Stats
| Metric | Value |
|---|---|
| Original Examples | 38,504 |
| Refusals Removed | 9,928 (26%) |
| Final Examples | 28,576 |
| Format | Messages (role + content) |
| Disk Size | ~996 MB |
| Source | angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k |
Source Models
The reasoning chains were distilled from:
- Claude Opus 4.6 — majority of examples
- Claude Opus 4.7 — latest reasoning capabilities
Categories include: coding, math, science, philosophy, creative writing, and more.
Format
{
"category": "coding",
"model": "claude-opus-4-6",
"messages": [
{"role": "system", "content": "You are a computer science tutor. Explain concepts clearly with examples."},
{"role": "user", "content": "How does a B-tree work?"},
{"role": "assistant", "content": "A B-tree is a self-balancing tree data structure...\\n\\nLet me walk through it step by step:\\n\\n1. Structure: ..."}
]
}
Filtering Process
Refusals were identified by scanning assistant responses for patterns:
I cannot/I can't/I'm unableas an AI/as a language modelI apologize/I'm sorryagainst my guidelines/not appropriate- And 7 more refusal patterns
9,928 examples (26%) removed — leaving 28,576 clean reasoning chains.
Why filter?
The base model for training (huihui-ai/Huihui-Qwen3.6-27B-abliterated) has been abliterated to remove refusals. Training on examples with refusals would reintroduce refusal behavior, defeating the purpose of the abliterated base.
Usage
from datasets import load_from_disk
dataset = load_from_disk("claude46_filtered")
print(len(dataset)) # 28576
From HuggingFace:
from datasets import load_dataset
dataset = load_dataset("hotdogs/claude46_filtered", split="train")
Training Status
This dataset is being used to train a LoRA adapter for Qwen3.6-27B-abliterated. Expected training stats:
- ~7,144 steps at effective batch size 4
- 1 epoch covers all 28,576 examples
- Target: 0% refusal reasoning model
Credits
- Original dataset: angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
- Filtering script: hotdogs/qwen3.6-27b-qlora-training