An LLM trained to write the sentence you always rewrite: the title

What it does

ChatGenTitle is a collection of LoRA adapters for LLaMA 7B and 13B trained on arXiv metadata—titles, abstracts, and other fields from over 2.2 million papers—to generate a research paper title from an abstract and a short prompt. The repository also includes a daily digest of newly crawled LLM papers from arXiv’s cs.AI, cs.CV, and cs.LG RSS feeds.

The interesting bit

The value is in the data curation: the authors used alpaca-lora to turn a general foundation model into a hyper-specific headline generator. Their comparison table against GPT-3.5, GPT-4, and ChatGLM suggests the fine-tuned model often mimics the concise style of real arXiv titles, whereas the generalist models tend to inflate the length and jargon.

Key highlights

• LoRA adapters for LLaMA-7B and LLaMA-13B, trained on both computer-science-specific and full arXiv corpora.
• Side-by-side comparisons with GPT-3.5, GPT-4, and ChatGLM show differing title-generation styles.
• Daily auto-crawled arXiv feeds for cs.AI, cs.CV, and cs.LG papers, published as a markdown digest.
• Released under a non-commercial CC BY-NC-SA 4.0 license; requires the original LLaMA base weights to function.
• Training logs and hardware details (A100 4×80GB) are openly shared.

Caveats

• The release table lists model sizes as -MB, suggesting the documentation has incomplete placeholders.
• Because Meta’s LLaMA license forbids commercial use, the adapters are strictly non-commercial and must be merged with the base model before inference.
• The README mixes Chinese and English and contains truncated sections, so some details—like exact training sample counts per model variant—are unclear.

Verdict

Academics who need a quick, stylistically consistent title suggestion will find this handy. If you already pay for GPT-4 and prefer its more verbose titles, or need a commercially deployable tool, look elsewhere.