AnmolNimmala0/agri-slm-corpus

AgriSLM Pre-Training Corpus

Domain: Agriculture (India-focused, English)
Version: 1.0
Total Documents: 266,691
Total Tokens: ~1.32 billion
File: corpus_filtered.jsonl (6.5 GB)

---

Purpose

Pre-training corpus for a 100–300M parameter agriculture-domain language model. Covers Indian and global agriculture: crops, soil science, irrigation, pest management, livestock, food security, government schemes, farmer Q&A, and research papers.

---

Document Schema

Every line in corpus_filtered.jsonl is a JSON object:

{
  "text": "Full article or document text",
  "source": "pubmed",
  "domain": "agriculture",
  "subdomain": "research_papers",
  "language": "en",
  "word_count": 3200,
  "url": "https://...",
  "title": "Article title"
}

---

Sources

Source	Docs	Share	Description
pubmed2	77,607	29.1%	PubMed open-access full-text papers (batch 2)
pubmed	76,759	28.8%	PubMed open-access full-text papers (batch 1)
europepmc	56,678	21.3%	Europe PMC open-access abstracts
europepmc2	12,524	4.7%	Europe PMC expansion queries
pubmed3	11,990	4.5%	PubMed batch 3 (mechanization, food chem, entomology)
core	11,868	4.5%	CORE open-access full-text papers
wikipedia_agri	8,317	3.1%	Wikipedia agriculture category articles
kcc_kaggle	6,072	2.3%	India KCC farmer helpline Q&A (government data)
extension	1,935	0.7%	Agricultural extension service articles
arxiv	2,917	1.1%	arXiv preprints (agriculture/food science)
irri	1,206	0.5%	IRRI rice research publications
fao	1,135	0.4%	FAO international agriculture documents
cimmyt	1,013	0.4%	CIMMYT maize/wheat research
icrisat	470	0.2%	ICRISAT dryland crops research
icar	556	0.2%	ICAR India national agriculture research
kisanvaani	548	0.2%	KisanVaani farmer Q&A
other	~5,000	~2%	TNAU, NABARD, universities, news portals

---

Quality Filters Applied

1. Language: English only (>60% ASCII characters)
2. Minimum length: 150 words (50 words for Q&A sources)
3. Maximum length: 10,000 words (removes conference abstract books)
4. Agriculture relevance: Must contain at least 1 agriculture keyword hit
5. Deduplication: MinHash LSH (80% Jaccard threshold, 128 permutations)

---

Loading the Dataset

Python (streaming, memory-efficient)

import jsonlines

with jsonlines.open("corpus_filtered.jsonl") as reader:
    for doc in reader:
        text = doc["text"]
        source = doc["source"]
        # process doc...

Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("json", data_files="corpus_filtered.jsonl", split="train")
# Access: ds[0]["text"], ds[0]["source"], etc.

Quick stats

import jsonlines
from collections import Counter

sources = Counter()
total_tokens = 0
with jsonlines.open("corpus_filtered.jsonl") as r:
    for doc in r:
        sources[doc["source"]] += 1
        total_tokens += int(doc.get("word_count", 0) / 0.75)
print(f"Total tokens: {total_tokens:,}")
print(sources.most_common(10))

---

Reproducing the Dataset

All collection and processing scripts are in scripts/:

scripts/
  collect_pubmed*.py       # PubMed via NCBI E-utilities
  collect_europepmc*.py    # Europe PMC REST API
  collect_wikipedia_agri.py
  collect_kcc_kaggle.py    # Kaggle KCC dataset
  collect_core.py          # CORE open-access API
  clean_dataset.py         # Quality filtering per source
  dedup_dataset.py         # MinHash LSH deduplication
  merge_corpus.py          # Shuffle and merge
  filter_corpus.py         # Post-processing filter

---

Known Limitations

• Skewed toward research papers (~88% from PubMed/EuropePMC abstracts/full-text)
• Limited Indian-language content — English only; Hindi/regional language agriculture content not included
• Government portal coverage is thin — ICAR, FAO, TNAU contribute <1% each due to scraping restrictions
• No real-time data — collection done June 2026; no live prices, weather, or news after that date

---

Contact

Built by Anmol Nimmala (amolnimmala@gmail.com) for Tribe3 agriculture SLM project.