RNAChat corpus: sentences paired with RNA
Each row pairs a natural-language request (Chinese or English) with an RNA, plus the structured condition the sentence was rendered from. Used to train RNAChat-v7.
Files
| path | content |
|---|---|
corpus/{train,val,test}.jsonl.gz |
1,298,904 / 104,812 / 105,703 rows |
eval/eval_{val,test}.jsonl.gz |
evaluation metadata (row, class, condition, sentence) used by the t04-t09 scripts |
eval/store_report.json |
statistics of the tokenised store |
store/ |
tokenised store the training code reads: seq_tokens.u8 (all sequences, 1 byte per base), seq_index.npz, rows_{train,val,test}.npz, text_*.bin (sentences) |
docs/ |
corpus list with sources (语料清单.md), schema design, data-preparation notes |
Train classes: mRNA 1,097,942; miRNA 49,768; ncRNA_other 92,453; tRNA 22,957; snRNA 16,829; snoRNA 14,924; rRNA 4,031.
Row fields
rna_class, species, condition (e.g. seed, family, arm, subtype, length_nt, gc, species, kind),
content (the sentence), lang, prompt_type, rna_id, seq_ref, split, weight, meta. corpus/*.jsonl.gz holds
no sequence: seq_ref is a pointer, and the sequences are in store/ (read with Store in tt_data.py; ids map
A/C/G/U via tt_common.ID_TO_BASE).
Sources and licences
RNAcentral, Rfam 15.1, miRBase (via RNAcentral), NCBI GeneRIF, EuropePMC abstracts (sampling), and transcript
annotations of human and other species. See docs/语料清单.md for URLs. Some sentences are LLM paraphrases
(Qwen2.5-7B-Instruct) of templates. Splits are made per class by the leakage unit described in docs/RNA通用Schema_v2.md
(for ncRNA, by Rfam family). Check each source's terms before making this repo public; license: other is a
placeholder.