j0no12/nero-titles-200k

Nero Chat Titles — Unified & Cleaned Dataset

A cleaned, deduplicated, and merged text-to-text dataset for chat title generation: given a conversation (or a user prompt), predict a short chat title.

Built from two public Hugging Face datasets, unified into one simple format, cleaned, deduplicated, and split deterministically for training.

📦 Dataset

Split	Examples
train	240,773
validation	13,376
test	13,377
total	267,526

Stat	Value
Total input tokens	42,055,688
Total target (title) tokens	1,908,298
Total training tokens	43,963,986 (~44M)
Avg input length	157.2 tokens (median 75)
Avg title length	7.1 tokens (median 7)

📐 Format

Each line of *.jsonl is valid JSON:

{
  "input": "essay on the topic lesson Receive in my life. 1800 words",
  "target": "Essay On Lesson Received",
  "source": "spitfire",
  "input_tokens": 13,
  "target_tokens": 7,
  "total_tokens": 20
}

• input — conversation / user prompt text (tokenized with t5-small)
• target — the chat title (label)
• source — which original dataset the example came from (spitfire or michionlion)
• input_tokens / target_tokens — t5-small token counts (precomputed)

🧹 Cleaning Pipeline

1. Dropped empty conversations, empty titles, and broken formatting
2. Dropped extremely short (< 50 chars) and extremely long (> 8,000 chars) conversations
3. Dropped invalid titles (non-alphanumeric only)
4. Removed 69,809 exact duplicates via MD5 fingerprinting of input||target
5. Normalized titles: trimmed whitespace, stripped stray punctuation, kept natural chat-title style (titles were not rewritten by hand)
6. Split 90 / 5 / 5 (train / validation / test) with a fixed seed (42) — deterministic and reproducible

🔗 Source Datasets & Attribution

This dataset is a merge of two community datasets, with a deeper lineage into SupraLabs' title datasets. All credit for the original data goes to their authors:

Source	License	Examples (raw)
Michionlion/chat-titles-english	CC-BY-4.0	~110k
spitfire4794/chat-titles-245k	—	~245k

The spitfire4794 dataset is itself a merge of two SupraLabs datasets:

Grandparent source	License
SupraLabs/supra-wild-titles-130k	ODC-BY
SupraLabs/chat-titles-filtered-115K	CC-BY-4.0

Final merged composition: ~60% from the spitfire4794/SupraLabs lineage, ~40% from Michionlion. If you use this dataset, please also cite the original source datasets above.

🎯 Intended Use

• Training small encoder-decoder models for chat title generation
• Any seq2seq task where the input is a short conversation/prompt and the output is a short title
• The dataset was prepared for training a ~29M parameter encoder-decoder Transformer (t5-small tokenizer, max_src_len=512, max_tgt_len=32)
• ~44M training tokens ≈ 1.75× the parameter count of a 25M model (Chinchilla-appropriate for small models)

⚠️ Notes

• English-only
• Single-turn prompts make up the majority; multi-turn conversations appear in the spitfire portion
• Token counts were computed with t5-small; other tokenizers will yield different numbers

🛠️ Reproduce the merge

# python 3.11+, requires: hf (huggingface-cli), pandas, pyarrow
# 1. Download sources
hf download Michionlion/chat-titles-english chat-titles-english-110k.jsonl --repo-type dataset
hf download spitfire4794/chat-titles-245k data/train-00000-of-00001.parquet --repo-type dataset
# 2. Merge, clean, dedupe, split (fixed seed 42), and export train/validation/test.jsonl

🏷️ Cite

@misc{chat_titles_unified_2026,
  title        = {Chat Titles -- Unified and Cleaned Dataset},
  author       = {j0no12},
  year         = {2026},
  note         = {Merged and cleaned from Michionlion/chat-titles-english and spitfire4794/chat-titles-245k}
}