EarthScience-Text-LLM-20K-90-10

This is a pure-text Earth-science corpus unified from three non-overlapping upstream datasets:

1. Ekimetrics/climateqa-ipcc-ipbes-reports-1.0: climate and IPCC/IPBES report chunks.
2. GeoGPT-Research-Project/GeoGPT-CoT-QA: geoscience question-answer reasoning.
3. gremlin97/RemoteSensingCorpus: remote-sensing and geospatial machine-learning text.

Files and Split

The previous preprocessing outputs were merged into a 23,098-record pool and then deterministically re-split into 90% train and 10% test.

Split	Records	Climate	Geoscience	Remote sensing
train	20,788	6,388	8,100	6,300
test	2,310	710	900	700
total	23,098	7,098	9,000	7,000

The split is source-stratified and uses seed 20260722. ClimateQA examples are grouped by source document; GeoGPT examples are grouped by DOI/title; remote-sensing chunks are grouped by exact text hash. A group is assigned wholly to one split. The resulting train/test group overlap and exact unit_hash overlap are both zero.

Unified Schema

Each JSONL line is one training unit:

{
  "id": "string",
  "split": "train | test",
  "domain": "climate | geoscience | remote_sensing",
  "source_dataset": "string",
  "source_split": "original preprocessing/source split",
  "task": "document_chunk | cot_qa",
  "text": "canonical language-model text",
  "instruction": "SFT prompt, empty for document chunks",
  "response": "SFT response, empty for document chunks",
  "messages": [],
  "metadata": {},
  "unit_hash": "sha1",
  "text_token_count": 123,
  "prompt_token_count": 12,
  "response_token_count": 111
}

Token Statistics

Token counts use tiktoken:cl100k_base, without chat-template or special-token overhead.

Split	Avg text	Avg prompt, all records	Avg response, all records	Avg prompt, non-empty	Avg response, non-empty
train	533.76	31.87	445.55	81.80	1,143.46
test	521.01	32.78	433.82	84.15	1,113.47
overall	532.49	31.96	444.37	82.03	1,140.46

There are 9,000 non-empty prompt/response records overall: 8,100 in train and 900 in test. The 14,098 document_chunk records intentionally have empty instruction and response; their content is stored in text. For this reason, both all-record and non-empty SFT averages are reported.

Preprocessing

• Normalized carriage returns, null bytes, repeated horizontal whitespace, and excessive blank lines.
• Removed too-short or malformed records and bounded unusually long GeoGPT examples during the original preprocessing stage.
• Converted all sources into one JSONL schema and computed canonical lowercase normalized-text SHA-1 hashes.
• Removed exact duplicates before this re-split.
• Merged the original preprocessing splits, then performed the deterministic source-stratified 90/10 group split described above.
• Recomputed prompt and response token counts for every record after splitting.
• Verified JSON parsing, required fields, split labels, record totals, group isolation, and exact-hash isolation.

Limitations and Licensing

This package combines upstream sources with different provenance and licensing terms. Consult each upstream dataset card and the per-record source_dataset/metadata fields before redistribution or commercial use. Long-form reasoning and report text may contain factual errors, extraction artifacts, citation fragments, or source-specific biases. The held-out split prevents exact/group leakage under the identifiers available in the source data, but it does not guarantee semantic decontamination against external benchmarks.