DragonData-Finance-Corpus

DragonData World Finance Database

Priced like infrastructure. Free like air. The institutional-grade bilingual (EN/中文) finance database for training and evaluating expert AI systems — collected, cleaned, deduplicated, tokenized, license-proven, timestamped. 3,700+ files and counting.

Start here: Browse the data ⬇ · Use it in 4 lines ⬇ · Live build feed ⬇

Browse the data

The table below is live — scroll, search, and read actual records right on this page. Full corpus: 134M-token training shards (data/) plus raw archives (raw/).

Live status

Next: shard 1,000 · backfill complete · CN-web veins at scale.

Why it is priceless

Commercial filings feeds bill tens of thousands per year. Millions of API filing reads cost hundreds of thousands. This database delivers the equivalent supply chain at zero, forever, to everyone.

Curation proof, not promises: duplicates measured and blocked, every figure traceable, every file checksummed. Honest scope: this data builds models that read and reason — never predict returns.

Quickstart

import numpy as np
from huggingface_hub import hf_hub_download

# 1. a training-ready shard (134M uint32 token ids, EOS-terminated documents)
fp = hf_hub_download("dragonlimited/DragonData-Finance-Corpus",
                     "data/train-00000-of-00001.bin", repo_type="dataset")
ids = np.fromfile(fp, dtype=np.uint32)

# 2. the matching tokenizer — exact round-trip, digits + CJK intact
from tokenizers import Tokenizer
tk = Tokenizer.from_file(hf_hub_download(
    "dragonlimited/DragonData-Finance-Corpus", "tokenizer/v1/tokenizer.json",
    repo_type="dataset"))
s = "Revenue HK$120,000 million; 净利润200亿"
assert tk.decode(tk.encode(s).ids) == s

# 3. raw records — 中文同样直接可用: {text, source, date, license}
import json
raw = hf_hub_download("dragonlimited/DragonData-Finance-Corpus",
                      "raw/sec_edgar/part-00000.jsonl", repo_type="dataset")
recs = [json.loads(l) for l in open(raw)]
print(recs[0].keys())

# 4. live counters without downloading data
info = json.load(open(hf_hub_download(
    "dragonlimited/DragonData-Finance-Corpus", "dataset_info.json",
    repo_type="dataset")))
print(info["num_shards"], "shards |", info["total_tokens"], "tokens |",
      round(info["total_bytes"] / 1e12, 3), "TB of", round(info["capacity_bytes"] / 1e12), "TB")

Gallery

How it is built

flowchart LR
    A[SEC / HKEX / Gov / News] --> B[License-tag + timestamp]
    B --> C[Raw JSONL archive]
    C --> D[Digit-exact tokenize]
    D --> E[134M-token shards]
    E --> F[HF mirror + state]
    F --> G[PROGRESS.md dashboard]
    G --> H[Monthly living + yearly frozen]

flowchart TD
    A[Box crashes / dies] --> B{What was lost?}
    B -->|In-progress buffer| C[Resume from spill + offsets]
    B -->|Whole box| D[New box reads HF state]
    C --> E[Continue at cursor — zero re-tokenizing]
    D --> E
    E --> F[Reconcile: resume at first shard gap]

Citation

dataset_info.json carries totals; metadata/build_log.json is the append-only changelog. Cite a release by commit hash.

@dataset{dragondata2026,
  title  = {DragonData World Finance Database},
  author = {DragonData},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/dragonlimited/DragonData-Finance-Corpus}
}

Companion eval CN-FinEval · this corpus is the single data home. No model repos exist or are planned on this account.