saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned

GPT-5.5 / Gemini 3.1 Pro / Grok 4 / Claude Fable & Mythos 5 / Qwen 3.7 Max Distillation Dataset (Cleaned v2)

Cleaned, deduplicated, and enhanced version of the Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset. 73 source datasets, 18M+ raw samples, reduced to 6.9M high-quality SFT-ready examples through exact dedup, heuristic filtering, and quality scoring.

Pre-formatted with a messages column — drop into any SFT pipeline (Axolotl, Unsloth, HF Trainer) with zero conversion.

v2 Improvements

Improvement	Detail
Chat format	messages column with [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
Exact dedup	65% reduction (19.8M to 6.9M)
Quality filtering	Removed identity mappings, one-word TRUE/FALSE/garbage responses, very short outputs
Quality scores	Per-row heuristic scoring (0-100) for downstream filtering
Provenance preserved	source and source_dataset columns retained

Cleaning Pipeline

Stage	Reduction	Description
Exact dedup (intra-category)	19.8M to 6.9M	SHA-256 hash match within each category
Exact dedup (cross-category)	—	Inter-category duplicate removal
Min response length	<50 chars dropped	Removes truncated/incomplete responses
Identity filter	instruction == response dropped	Removes no-op training pairs
TRUE/FALSE filter	One-word verdicts dropped	Removes template garbage
Quality threshold	score < 40 dropped	Heuristic filter for remaining noise

Dataset Statistics

Category	Rows	Size	Avg Quality
coding	6,314,576	356 MB	51.9
cybersecurity	396,206	386 MB	55.1
instruction	52,238	16 MB	53.4
distilled	65,297	13 MB	59.9
index	25,349	3 MB	52.5
humanities	6,248	239 KB	51.1
science	3,158	338 KB	51.2
applied	663	70 KB	52.6
Total	6,863,735	774 MB	52.4

Top Sources

Source	Rows
vibe_instruct_v2	5,000,000
vibe_coding	1,100,000
high_quality_cybersecurity	269,635
fable5_agentic_sft	157,487
fenrir_v2_1	98,711
royal_ghost_501k	41,205
alpaca	39,685
gemini35	25,000
gemini_pro32	25,000
species_25k	25,000
savani_cyber_attack	13,883
precinct6_cybersecurity	13,119
dolly	12,553
kimi_coding	8,707
claude_mythos	5,315
sports_25k	5,173
claude_opus_48_distill	5,000
claude_opus_48_max_thinking	4,982
legend_python	4,000
(20+ more)

Data Fields

Column	Type	Description
messages	List[Struct{role: str, content: str}]	Chat-format conversation (SFT-ready)
instruction	str	Raw user instruction
response	str	Raw assistant response
source	str	Source split name (e.g. vibe_instruct_v2)
source_dataset	str	Original Hugging Face dataset path
category	str	Content category
quality_score	float64	Heuristic quality score (0-100)

Files

All files in v2/ directory:

File	Description
full_clean.parquet	Combined dataset, single file (812 MB)
coding.parquet	Coding SFT data (largest split)
cybersecurity.parquet	Cybersecurity SFT data
instruction.parquet	General instruction data (Alpaca, Dolly)
distilled.parquet	Distilled responses (Gemini, Claude, Fable, Mythos)
index.parquet	Index corpus
humanities.parquet	Humanities domain
science.parquet	Science domain
applied.parquet	Applied science
stats.json	Processing statistics

Usage

from datasets import load_dataset

ds = load_dataset(
    "saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned",
    split="v2"
)

print(f"Total examples: {len(ds)}")
print(ds[0]["messages"])

# Filter by quality score
ds = ds.filter(lambda x: x["quality_score"] >= 60)

# Filter by category
coding_ds = ds.filter(lambda x: x["category"] == "coding")

# Filter out known low-quality sources
ds = ds.filter(lambda x: x["source"] not in {
    "genesis_v11", "god_coder", "python_god_coder"
})

# Axolotl config
datasets:
  - path: saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned
    type: chat_template
    field_messages: messages
    split: v2

# Unsloth config
dataset: saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned
dataset_args:
  split: v2

Citation

@misc{distillation-cleaned-v2,
  author = {Sai Dutta},
  title = {GPT-5.5 / Gemini 3.1 Pro / Grok 4 / Claude Fable and Mythos 5 / Qwen 3.7 Max Distillation Dataset (Cleaned v2)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned}}
}