j0no12/Nero-Titles

Nero Titles

A 727,319-row, English-dominant chat-title training corpus with published source counts, safety filtering, exact-input deduplication, and TitleBench exclusion.

Quick start

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="hf://datasets/j0no12/Nero-Titles/data.jsonl",
    split="train",
)

print(len(dataset))
# 727319
print(dataset[0])

Dataset summary

Property	Value
Final rows	727,319
Raw rows ingested	823,395
Synthetic sources before filtering/deduplication	569,246
Real sources before filtering/deduplication	254,149
Unsafe rows removed	2,725
TitleBench-overlap rows removed	50,974
Duplicate-input rows removed	42,375
Empty rows removed	2
Split	Training collection only

Counts come from the released stats.json. They replace smaller figures in the earlier card that did not match the published artifact.

Schema

Each JSONL row uses:

Field	Type	Description
input	string	Chat prompt or conversation text
title	string	Short target title
source	string	Provenance label

Training example formatter:

def to_pair(row):
    return {"source_text": row["input"], "target_text": row["title"]}

Some downstream pipelines may rename title to target; the released JSONL field is title.

Source lineage

The released statistics describe raw ingested rows before the global removal steps:

source label	Raw rows	Description
synthetic_teacher	370,367	Teacher-generated titles over LMSYS-derived inputs
synthetic_bananamind	198,879	BananaMind-teacher titles reconstructed from gated LMSYS rows
real_unified_train	240,773	Human-title training rows from j0no12/chat-titles-unified
real_unified_val	13,376	Human-title validation rows from the same unified source
Total raw ingested	823,395	Before filtering and deduplication

The real unified source derives from spitfire4794/chat-titles-245k and Michionlion/chat-titles-english. Synthetic inputs derive from locally accessed lmsys/lmsys-chat-1m; raw gated LMSYS text should remain subject to that dataset's access terms.

Construction story

1. Ingested four source streams totaling 823,395 rows.
2. Removed 2 empty rows.
3. Removed 2,725 rows matching the project's unsafe-content indicators.
4. Excluded 50,974 rows whose normalized input matched the full reconstructed TitleBench input set.
5. Deduplicated globally on normalized input, removing 42,375 additional rows.
6. Serialized the remaining 727,319 rows as data.jsonl.

The arithmetic is auditable: 823,395 - 2 - 2,725 - 50,974 - 42,375 = 727,319.

Quality and integrity gates

• Benchmark exclusion: normalized inputs matching all 21,448 TitleBench examples, including locally reconstructed gated examples, were removed.
• Safety filter: the construction pipeline removed rows matching its child-safety, sexualized-minor, or explicit-content indicators.
• Exact-input deduplication: only one example per normalized input was retained under the pipeline's source-priority rule.
• Schema normalization: every released row uses string input, title, and source fields.
• Published accounting: stats.json records raw-source and removal counts.

These gates are pipeline-level checks, not a substitute for an independent human audit.

Contamination screening

The release specifically screens normalized-input overlap with j0no12/TitleBench. This supports a held-out TitleBench evaluation for models trained only on this artifact. It does not guarantee removal of paraphrases, semantic near-duplicates, matching targets paired with different inputs, or benchmark material already learned during external pretraining.

Intended use

• Training or fine-tuning encoder-decoder and decoder-only chat-title models.
• Comparing synthetic-title and human-title mixtures.
• Research on compact, task-specialized generation models.

Do not use the corpus as a general conversational assistant dataset or as evidence that a generated title is factually correct or safe.

Known limitations

• English-dominant inputs and short titles, typically a few words.
• A large portion of the raw mixture is synthetic and inherits teacher preferences and mistakes.
• Safety filtering is keyword/indicator based and may have false positives or false negatives.
• Exact normalized-input deduplication does not remove all semantic duplication.
• No separate development or test split is shipped in this repository.
• No published human-quality study, PII audit, or broad benchmark-decontamination report accompanies the release.
• Repository-level MIT metadata does not override licenses, access restrictions, attribution requirements, or model-output terms attached to upstream sources.

Files and integrity

File	Size	Integrity identifier
data.jsonl	340,630,278 bytes	SHA-256 97ff5a0c3c5ad27023c767004453a20be84de662ff7b465e4cecbb0e5caad153
stats.json	311 bytes	Git blob 892aa235057ae05490a370774103c346f4b8f457
Repository revision documented by this card update	—	03697a3598e983fb475596dca8bdeae02b88dd01

Citation

@dataset{j0no12_nero_titles_2026,
  author    = {j0no12},
  title     = {Nero Titles},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/j0no12/Nero-Titles}
}

Please also credit the upstream datasets and teacher models relevant to the subset you use.

License

Mixed / other. Upstream dataset terms, gated LMSYS access requirements, and teacher-model output terms remain applicable; this card does not replace them with a blanket license or grant additional rights to upstream material.