jsonl to binidx tool

This repository can serve for dataset preparation of RWKV model,

To speedup the tokenization, please install the fast RWKV Tokenizer written in Rust.

pip install pyrwkv-tokenizer

The multilingual rwkv-6-world models use a new tokenizer rwkv_vocab_v20230424.txt.

python tools/preprocess_data.py --input ./sample.jsonl --output-prefix ./data/sample --vocab ./rwkv_vocab_v20230424.txt --dataset-impl mmap --tokenizer-type RWKVTokenizer --append-eod

The jsonl format sample (one line for each document):

{"text": "This is the first document."}
{"text": "Hello\nWorld"}
{"text": "1+1=2\n1+2=3\n2+2=4"}

generated by code like this:

ss = json.dumps({"meta": meta, "text": text}, ensure_ascii=False)
out.write(ss + "\n")