---
license: apache-2.0
language:
- it
- en
size_categories: - n<1K
task_categories: - text-generation
- question-answering
tags: - code
- coding
- uncensored
- distilled
- deepseek
- italian
CrimeOpus 4.7 — Distilled Coding Dataset
Dataset di fine-tuning per CrimeOpus 4.7-v2 (LoRA training).
Sources
| Source | Count | Type |
|---|---|---|
| DeepSeek-Chat distillation | 157 | Multi-domain coding/reasoning |
| Git commit-diff (CrimeCode-IDE) | 91 | Real codebase patterns |
| Uncensored seed (toxic-dpo + orpo-mix) | 179 | Refusal-free helpfulness |
| Total | 427 |
Format
ChatML messages array:
{
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
]
}
Stats
- Avg user message: ~281 chars
- Avg assistant response: ~2745 chars
- Total tokens (est): ~323k
- Languages: Italian + English + code
- Refusals filtered: yes (zero detected)
Use
from datasets import load_dataset
ds = load_dataset("JollyFraud/crimeopus-distill-v2", split="train")
Used for QLoRA fine-tuning of huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated.