anon_reasoning_v5
anon_reasoning_v5 is a curated, high-information-density SFT dataset designed to fine-tune language models with an authentic imageboard / anon conversational persona without degrading core reasoning, coding, and mathematical capabilities.
What Makes v5 Different?
Previous iterations (v3.2 and v4) suffered from severe slop dilution: over 80% of those datasets consisted of sub-50-character thread replies ("fuck no", "based", "cope"), which caused models to develop aggressive early-truncation (EOS) bias and catastrophic forgetting on multi-step reasoning.
v5 re-balances the signal-to-noise ratio:
- Zero Sub-50 Character Throwaways: Eliminated ~13,700 trivial one-word replies. Overall median assistant length increased from 90 characters to 383 characters (avg 776 characters).
- Preserved Pure Chain-of-Thought: Features 2,583 verified
<think>CoT reasoning examples (supra_cot), ensuring the model maintains deep step-by-step thinking capabilities. - High-Quality Instruction Following: Retains 4,458 multi-domain instructional chat examples (
fable_chat, median length 1,542 chars). - Verified Greentext Corpus: 2,887 verified, multi-line greentext stories (
cesspool_greentext, median length 455 chars) filtered for proper story format (>). - No Poisoned Reasoning Mappings: Stripped corrupted "stitched" pairings where serious logic puzzles were mapped to unrelated forum insults.
- Deduplicated Prompts: Removed redundant duplicate user turns to prevent mode collapse.
Dataset Composition
| Type | Rows | Share (%) | Median Asst Length | Description |
|---|---|---|---|---|
cesspool_chat |
11,505 | 52.6% | 169 chars | Curated conversational persona, slang, and banter (length 80–3000 chars, deduplicated) |
fable_chat |
4,458 | 20.4% | 1,542 chars | Multi-turn & single-turn high-quality instruction following (coding, STEM, creative) |
cesspool_greentext |
2,887 | 13.2% | 455 chars | Authentic multi-line greentext stories (>be me...) |
supra_cot |
2,583 | 11.8% | 840 chars | Pure Chain-of-Thought reasoning with explicit <think> and </think> tags |
topical_stitched_fable_chat |
316 | 1.5% | 1,067 chars | Long-form verified instructions |
cesspool_4claw |
105 | 0.5% | 573 chars | Curated agentic/persona interactions |
| Total | 21,854 | 100% | 383 chars | Optimal Reasoning-Persona Balance |
Format
The dataset uses the standard ChatML / OpenAI message format:
{
"messages": [
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
],
"type": "supra_cot"
}
A compatible chat_template.jinja (ChatML) is included in the repository.
Disclaimers & Licenses
Built from curated subsets of:
SupraLabs/Supra-Reasoning-CoT-Dataset-v2saidutta69/fable-5-premium-v2v2ray/4chanscratchyone/anon-greentextDavidTKeane/4claw-ai-agent-dataset
Contains informal, offensive, uncensored, and imageboard-style text by design. Each upstream source maintains its respective license and terms.