Michionlion/chat-titles-english

Chat Titles English 110K

Chat Titles English 110K is an English prompt-to-title dataset for training and evaluating models that generate concise, retrievable conversation titles from a user's first message.

The release contains 110,325 JSONL rows. Each row includes a deterministic UUIDv5 identifier, the user prompt, a single-line title of at most 50 characters, and zero or more heuristic quality flags.

The dataset is broad and general-purpose, containing a significant amount of synthesized examples as well as some realistic examples.

Dataset summary

Property	Value
Rows	110,325
File size	79.3 MiB
Language	Predominantly English
Official splits	One train split
Maximum user length	6,144 characters
Maximum title length	50 characters
Median user length	266 characters
Median title length	26 characters
Median title length	4 words
Unique exact titles	85,336
Rows with one or more flags	40,717
Rows with an empty flags array	69,608
Exact normalized duplicate prompts	0
SHA-256	30e32555c79c00abf8b2c9b107e4b7a7edbf68dd67eb2cbe42d5595efb964ac8

Intended task

Given a user message, generate a short title that helps the user find the conversation later.

Suitable uses include:

• Supervised fine-tuning for thread or chat title generation
• Title-generation benchmarking
• Pairwise ranking or preference evaluation of candidate titles
• Training compact title-generation models
• Research on concise intent summarization

The dataset is not intended as a general question-answering corpus, a code-generation corpus, or a safety benchmark.

Data format

Each line is a JSON object with the following fields:

Field	Type	Description
guid	string	Deterministic RFC 4122 UUIDv5 identifier
user	string	User prompt or first conversation message
title	string	Curated single-line title, at most 50 characters
flags	list[string]	Heuristic quality or distribution risks; empty when no current issue is flagged

Example:

{
  "guid": "e1e45426-8f39-59d7-b269-8d257a4ed595",
  "user": "Could you help me remove the click that appears when this oscillator restarts? I am new to real-time audio code, so I may be resetting phase in the wrong place.\n\n```cpp\nif (gateRising) {\n    phase = 0.0f;\n}\nout = std::sin(phase) * env;\nphase += frequency / sampleRate;\n```",
  "title": "Oscillator restart click",
  "flags": []
}

The final schema intentionally does not include a row-level source field. Source composition is documented below at the dataset level.

Loading the dataset

From a local JSONL file:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="chat-titles-english-110k.jsonl",
    split="train",
)

From the Hub:

from datasets import load_dataset

dataset = load_dataset(
    "Michionlion/chat-titles-english-110k",
    split="train",
)

To use only rows without current heuristic flags:

clean_subset = dataset.filter(lambda row: len(row["flags"]) == 0)

Source composition and attribution

Component	Final rows	Share	Source license/status
SupraLabs/chat-titles-unfiltered-150K-derived component	92,405	83.76%	CC BY 4.0
kth8/title-generation-25000x, english split	17,040	15.45%	Apache 2.0
software persona set	500	0.45%	Generated for this project; CC-by-4.0
kth8/title-generation-25000x, English-retained rows from multilingual	338	0.31%	Apache 2.0
kth8/title-generation-25000x, supplementary split	42	0.04%	Apache 2.0
Total	110,325	100%	Mixed provenance

SupraLabs-derived component

The largest component began with 149,743 rows from SupraLabs/chat-titles-unfiltered-150K and was reduced to 92,405 rows through language filtering, validity checks, artifact removal, title review, and deduplication.

The upstream SupraLabs dataset card states that:

• User prompts originated from sam-mosaic/orca-gpt4-chatml.
• Titles were generated with Qwen3.6-35B-A3B.
• Additional review used Qwen3.5-9B-Q8.
• The released SupraLabs dataset is licensed under CC BY 4.0.

OpenCode title-generation component

The files named english.jsonl, multilingual.jsonl, and supplementary.jsonl correspond to the three splits of kth8/title-generation-25000x, which is tagged Apache 2.0.

The original files included text, title, uuid, and, in some splits, model reasoning. This release retains only the prompt as user, and the title. The original reasoning traces are not included, and the uuid was regenerated.

Related title-generator model cards include:

• kth8/gemma-3-270m-it-OpenCode-Title-Generator
• kth8/gemma-3-1b-it-OpenCode-Title-Generator

Software persona component

The 500 persona component rows were generated with OpenAI GPT-5.6 Sol from 500 distinct software-development personas spanning 50 roles. The generation process varied seniority, tone, verbosity, casing, layout, politeness, punctuation style, technical context, and the type of request being made.

These rows were created specifically to increase the realism and stylistic diversity of coding and technical prompts. The final dataset retains the prompt and title but not the persona metadata or generation traces.

The persona rows should be treated as synthetic data. Although they were reviewed and filtered as part of the broader curation pipeline, they may still reflect model-generated patterns, stylistic regularities, or factual artifacts.

Curation pipeline

Dataset combination, filtering, quality review, deduplication, schema normalization, report generation, and parts of the manual relabeling workflow were performed with the assistance of OpenAI GPT-5.6 Sol. The model was used as a curation and analysis tool under human direction; final inclusion criteria, thresholds, and release decisions were selected by me.

The automated and model-assisted steps were supplemented with deterministic validation, explicit threshold checks, manual inspection of edge cases, and manual rewriting of selected titles. Use of GPT-5.6 Sol does not imply that OpenAI reviewed, endorsed, or verified this dataset.

1. Schema normalization

Source fields were normalized to:

guid, user, title, flags

All final GUIDs were regenerated as deterministic RFC 4122 UUIDv5 values using:

• Namespace UUID: 141c74db-914c-52eb-95ac-cf989289bef3
• Name pattern: row:000001, row:000002, and so on

The identifiers are stable for this exact ordered release. Reordering the dataset and regenerating identifiers would produce different GUIDs.

2. Language filtering

Language filtering used two stages:

1. Conservative Unicode screening

   • Removed clear non-English alphabetic scripts and accented language characters.
   • Ignored fenced code, inline code, URLs, and LaTeX-style math.
   • Retained punctuation, operators, code-like symbols, mathematical Unicode letters, and isolated Greek variables.

2. Code-aware language classification

   • Removed code and math spans before classification.
   • Used FastText lid.176 on prose-like content.
   • Non-English candidates at approximately 70% or greater confidence were independently checked with Lingua.
   • Ambiguous short ASCII-only cases were reviewed by OpenAI GPT-5.6 Sol and retained when uncertain.

This process aims to remove predominantly non-English prose without discarding prompts that contain non-English identifiers, technical notation, or code.

3. Length and validity checks

Final constraints:

• user must be non-empty and no longer than 6,144 characters.
• title must be a single line and no longer than 50 characters.
• Rows with mojibake, replacement characters, invalid controls, Unicode noncharacters, or similar encoding corruption were removed.
• Over-length or low-quality titles were reviewed and rewritten by OpenAI GPT-5.6 Sol.

4. Low-quality artifact removal

High-confidence synthetic or low-value prompt families were removed, including:

• NLI and entailment benchmark scaffolding
• Paraphrase and question-similarity classification
• Sentiment and toxicity classification artifacts
• Synthetic task wrappers
• Question-generation benchmarks
• Word-manipulation exercises
• Implausible or counterfactual generation tasks
• Explicit title-generation leakage
• The repeated Movie Review Sentiment family

5. Deduplication

Prompts were normalized with Unicode NFKC normalization, case folding, punctuation removal, and whitespace normalization.

• Exact normalized duplicates were removed.
• Near duplicates were removed only when normalized RapidFuzz Indel similarity was at least 95%.

Approximate template-family repetition can still remain below the 95% threshold.

Canonical title policy

The following prompt was used as the source of truth for manual relabeling and title validation; it is the system prompt used by OpenCode for title classification.

Canonical title-generation system prompt

You are a title generator. You output ONLY a thread title. Nothing else.

<task>
Generate a brief title that would help the user find this conversation later.

Follow all rules in <rules>
Use the <examples> so you know what a good title looks like.
Your output must be:
- A single line
- ≤50 characters
- No explanations
</task>

<rules>
- you MUST use the same language as the user message you are summarizing
- Title must be grammatically correct and read naturally - no word salad
- Never include tool names in the title (e.g. "read tool", "bash tool", "edit tool")
- Focus on the main topic or question the user needs to retrieve
- Vary your phrasing - avoid repetitive patterns like always starting with "Analyzing"
- When a file is mentioned, focus on WHAT the user wants to do WITH the file, not just that they shared it
- Keep exact: technical terms, numbers, filenames, HTTP codes
- Remove: the, this, my, a, an
- Never assume tech stack
- Never use tools
- NEVER respond to questions, just generate a title for the conversation
- The title should NEVER include "summarizing" or "generating" when generating a title
- DO NOT SAY YOU CANNOT GENERATE A TITLE OR COMPLAIN ABOUT THE INPUT
- Always output something meaningful, even if the input is minimal.
- If the user message is short or conversational (e.g. "hello", "lol", "what's up", "hey"):
  → create a title that reflects the user's tone or intent (such as Greeting, Quick check-in, Light chat, Intro message, etc.)
</rules>

<examples>
"debug 500 errors in production" → Debugging production 500 errors
"refactor user service" → Refactoring user service
"why is app.js failing" → app.js failure investigation
"implement rate limiting" → Rate limiting implementation
"how do I connect postgres to my API" → Postgres API connection
"best practices for React hooks" → React hooks best practices
"@src/auth.ts can you add refresh token support" → Auth refresh token support
"@utils/parser.ts this is broken" → Parser bug fix
"look at @config.json" → Config review
"@App.tsx add dark mode toggle" → Dark mode toggle in App
</examples>

Quality flags

Flags identify rows that may still be useful but carry a known quality, leakage, or distribution risk. They are review signals, not ground-truth labels, and they are not mutually exclusive.

Flag	Rows	Share	Meaning
answer_leakage_or_template	26,680	24.18%	Contains answer/output scaffolding or template-like benchmark phrasing
multiple_choice_benchmark	13,015	11.80%	Contains explicit answer choices or multiple-choice benchmark structure
translation_benchmark	3,578	3.24%	Translation-style task with benchmark-like formatting
news_article_transformation	2,074	1.88%	News summarization, continuation, rewriting, or article transformation
minimal_prompt	34	0.03%	Very short prompt retained because it still has plausible conversational value
title_generation_leakage	2	<0.01%	Prompt explicitly asks for a title, risking target-task leakage

Because flags overlap, their counts sum to more than the number of flagged rows.

Top flag combinations:

Flags	Rows
No flags	69,608
answer_leakage_or_template	22,018
multiple_choice_benchmark	8,352
Both answer/template and multiple-choice	4,659
translation_benchmark	3,571
news_article_transformation	2,074

Statistical profile

Prompt length

Statistic	Characters
Mean	630.5
Median	266
90th percentile	1,909
95th percentile	2,306
99th percentile	4,073.8
Maximum	6,144

Title length

Statistic	Characters
Mean	26.82
Median	26
90th percentile	37
95th percentile	42
99th percentile	48
Maximum	50

Titles contain a median of 4 words and a mean of 3.79 words.

Title repetition

• Exact unique titles: 85,336
• Repeated-title groups: 4,486
• Rows belonging to a repeated-title group: 29,475
• Additional rows beyond one representative per repeated title: 24,989

Repeated titles do not necessarily imply duplicate prompts. Many distinct prompts legitimately share concise titles, while some repeated generic titles reflect benchmark-family concentration.

Coding and code-content coverage

Measure	Rows	Share
Clearly coding-related	2,060	1.87%
Broader coding or infrastructure-related	2,531	2.29%
Contains actual code or executable-style snippets	461	0.42%

The 461 code-containing rows were categorized as:

• 265 with fenced code blocks
• 102 with unfenced multiline code
• 7 with other unfenced code snippets
• 87 with inline code expressions or commands

This release is therefore a general thread-title dataset with limited coding coverage, not a coding-majority corpus. Users targeting software-development title generation should add or upsample high-quality coding data.

Data splits

Only one train split is provided.

For rigorous evaluation, do not create a purely random split. Prompt templates and benchmark families can remain similar even after 95% deduplication. A stronger procedure is to cluster prompts by semantic or template similarity and assign entire clusters to train, validation, or test.

Known limitations

• A substantial minority of rows retain benchmark-like formatting and are marked with flags.
• The flagging system is heuristic and may contain false positives and false negatives.
• Near-duplicate detection is high precision at 95%, but it is not proof that no lower-similarity template duplicates remain.
• Repeated generic titles and highly represented benchmark families may bias models toward broad titles.
• Coding coverage is small relative to the full corpus.
• The dataset is English-focused and should not be used to measure broad multilingual title generation.
• No dedicated full-corpus PII audit or safety-classification pass was performed for this release.
• Source-generated titles may contain factual, stylistic, or alignment errors despite curation.
• UUIDs are stable for this release order, not content-addressed identifiers.

Privacy and safety

The dataset contains prompts rather than assistant answers, but prompts can still include personal information, unsafe content, or copyrighted text. Users should perform their own privacy, safety, and licensing review before training or deploying a production model.

Do not use this dataset to identify individuals or infer sensitive personal data.

Licensing

Important source terms include:

• SupraLabs/chat-titles-unfiltered-150K: CC BY 4.0; attribution should be retained.
• kth8/title-generation-25000x: Apache 2.0.

The license applied to the compilation does not erase or replace the terms of source components. Downstream users are responsible for satisfying all applicable source licenses and notices. This section is informational and not legal advice.

Attribution

When using this dataset, please cite the published repository and retain attribution to the upstream datasets.

AI assistance disclosure

Suggested citation for this release:

@dataset{chat_titles_english_110k_2026,
  title     = {Chat Titles English 110K},
  author    = {Michionlion},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Michionlion/chat-titles-english-110k}
}

SupraLabs attribution:

@dataset{supralabs_chat_titles_unfiltered_150k_2026,
  author       = {QyrouNnet-AI},
  title        = {Chat Titles Unfiltered 150K},
  year         = {2026},
  organization = {SupraLabs},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/SupraLabs/chat-titles-unfiltered-150K}
}

Additional source links:

• SupraLabs/chat-titles-unfiltered-150K
• kth8/title-generation-25000x

Acknowledgements

Thanks to the maintainers and contributors of the upstream SupraLabs and kth8 datasets and model repositories, and to the broader open-source title-generation community.

I used OpenAI GPT-5.6 Sol to assist with dataset combination, language filtering, quality auditing, deduplication, title review, documentation, and release preparation. All model-assisted work was directed and reviewed by me.