KrishiGyan (কৃষিজ্ঞান)

A Bengali question–answer dataset for Bangladeshi agriculture, with an explicit chain-of-thought reasoning trace on every entry.

Bangladesh has decades of agricultural research sitting in printed pamphlets, extension leaflets and encyclopedia entries. Very little of it is machine-readable, and almost none of it exists in a form a language model can be trained on. KrishiGyan is an attempt to close part of that gap: 5,529 Bengali Q&A pairs grounded in publications from BARI, BRRI, DAE, SRDI, Banglapedia and other Bangladeshi institutions, each one carrying the reasoning steps that lead from question to answer.

To our knowledge it is the first publicly released Bengali agricultural dataset with explicit reasoning chains.

Entries	5,529
Language	Bengali (bn)
Reasoning traces	5,529 (100%)
Categories	8
Source references	441
Format	JSONL (auto-converted to Parquet on the Hub)
License	CC BY 4.0
Human-validated accuracy	96.5% on a 200-entry stratified sample
Paper	PEEIACON 2026 (IEEE), to appear
Version	v1.0 — the release accompanying the paper

---

Quick start

from datasets import load_dataset

ds = load_dataset("AROY76/KrishiGyan")          # 5,529 entries, single train split
print(ds["train"][0]["instruction"])

Load a single category:

ds = load_dataset(
    "AROY76/KrishiGyan",
    data_files="data/soil_nutrient_management.jsonl",
)

Load the human-validation audit sample together with its reviewer scores:

sample = load_dataset("AROY76/KrishiGyan", "human_validation")         # 200 entries
scores = load_dataset("AROY76/KrishiGyan", "human_validation_scores")  # 200 rating rows

Format an entry for reasoning-style supervised fine-tuning:

def to_sft(row):
    target = f"<think>\n{row['thinking']}\n</think>\n\n{row['response']}"
    return {"prompt": row["instruction"], "completion": target}

---

Dataset structure

Files

data/
├── crop_production_protection.jsonl        1,387 entries
├── soil_nutrient_management.jsonl          1,164
├── technology_innovation.jsonl               720
├── water_irrigation_climate.jsonl            505
├── livestock_fisheries.jsonl                 477
├── research_education_extension.jsonl        456
├── organic_sustainable_agriculture.jsonl     416
└── economics_policy_social_science.jsonl     404
validation/
├── human_validation_sample.jsonl             200 entries reviewed by hand
└── human_validation_results.csv              the reviewer's scores for those 200
krishigyan_unsloth_datasets.ipynb             dataset prep / fine-tuning notebook

The eight category files together form the default config. They are disjoint — every entry appears exactly once, and all 5,529 id values are unique.

Fields

Field	Type	Description
id	string	Unique identifier, e.g. ag_001, soil_bangla_562
instruction	string	The question, in Bengali
thinking	string	Step-by-step reasoning trace — the logical path an agriculturist would follow, not a summary of the answer
response	string	The answer, in Bengali, written in an instructional register
metadata	object	Provenance and classification, described below

Metadata

Key	Coverage	Description
category	5,529	One of the eight categories
difficulty	5,529	easy / medium / hard, plus four legacy aliases (see note below)
season	5,529	all_year / winter / monsoon / summer (see note below)
topic	5,529	Topic label — a crop or livestock name (rice, jute, fish) or a subject label, in English or Bengali. 1,779 distinct values
source	5,529	Named source reference (441 distinct values)
source_type	5,529	Provenance class, see table below

All six keys are present on all 5,529 entries with no null or placeholder values, so the eight files concatenate without a schema cast.

Example

{
  "id": "agri_guide_013",
  "instruction": "আউশ ধান কত দিনে পাকে এবং এটি বর্ষাকালীন ফসল হিসেবে কেন উপযুক্ত?",
  "thinking": "প্রশ্নটি আউশের মেয়াদ ও বর্ষায় উপযুক্ততা সম্পর্কে।\n\nআউশের মেয়াদ:\n- ৯০-১২০ দিন (সংক্ষিপ্ত জাত)\n\nবর্ষায় কেন উপযুক্ত:\n- বর্ষার শুরুতেই বপন করা যায়\n- বন্যা আসার আগেই কাটা যায় ...",
  "response": "আউশ ধানের মেয়াদ ও বর্ষায় উপযুক্ততা:\n\n**মেয়াদ:**\n- জীবনকাল: ≈৯০–১২০ দিন\n- এটি সংক্ষিপ্ত-মেয়াদী (Short duration) জাত\n\n**বর্ষাকালে কেন উপযুক্ত:**\n\n১. **আগাম ফসল:** ...",
  "metadata": {
    "category": "Crop Production & Protection",
    "difficulty": "medium",
    "season": "monsoon",
    "topic": "aus_rice",
    "source": "Expert-curated synthetic data on general agriculture based on DAE extension materials",
    "source_type": "synthetic"
  }
}

English translation of the example

Question: In how many days does Aus rice mature, and why is it well suited as a monsoon crop?

Reasoning: The question concerns the duration of Aus and its suitability during the monsoon. Duration: 90–120 days for short varieties. Why it suits the monsoon: it can be sown at the start of the rains; it can be harvested before flooding arrives; the land is freed before Aman is transplanted; it needs little irrigation because rainfall covers it. Benefits: early income; the same plot can carry a later Aman or Rabi crop; it is safer in flood-prone areas.

Answer: Duration — a life cycle of roughly 90–120 days, a short-duration variety. Why it suits the monsoon: (1) Early crop — sown at the start of the rains and harvested before flooding; (2) Rain-fed — needs little irrigation; (3) Frees the land — Aman or a Rabi crop can follow on the same plot; (4) Lower risk — in flood-prone areas the early harvest avoids losses. Sowing is usually April–May, harvest July–August.

---

Statistics

Categories

Category	Entries	Share
Crop Production & Protection	1,387	25.09%
Soil & Nutrient Management	1,164	21.05%
Technology & Innovation	720	13.02%
Water, Irrigation & Climate	505	9.13%
Livestock & Fisheries	477	8.63%
Research, Education & Extension	456	8.25%
Organic & Sustainable Agriculture	416	7.52%
Economics, Policy & Social Science	404	7.31%
Total	5,529	100%

Crop production and soil management together account for roughly 46% of the dataset, which mirrors both their weight in Bangladeshi agriculture and the volume of source material available for them.

Difficulty

Difficulty	Entries	Share
Easy — factual recall (e.g. recommended planting time for Boro rice)	1,808	32.70%
Medium — multi-factor application (e.g. adjusting fertiliser rates by soil type)	2,336	42.25%
Hard — comparative or analytical reasoning (e.g. trade-offs between shrimp aquaculture and rice in the saline coastal zone)	1,174	21.23%
Legacy aliases	211	3.82%

The 211 legacy aliases are earlier label names that survived consolidation. Map them before training:

ALIAS = {"basic": "easy", "intermediate": "medium", "advanced": "hard", "high": "hard"}

Counts are intermediate (130), advanced (59), basic (21), high (1). After mapping: easy 1,829 (33.08%), medium 2,466 (44.60%), hard 1,234 (22.32%).

Provenance

source_type	Entries	Share
synthetic	2,094	37.87%
technical_guide	1,515	27.40%
encyclopedia	1,192	21.56%
government_publication	418	7.56%
government_research	237	4.29%
field_knowledge	49	0.89%
statistical_publication	24	0.43%

The season field

season	Entries	Share
all_year	3,914	70.79%
winter	709	12.82%
monsoon	692	12.52%
summer	214	3.87%

season was normalised to this closed four-value vocabulary before release. The raw annotations held 110 distinct values that mixed English snake_case (rabi, boro, kharif), Bengali script (সারা বছর, বোরো, আমন), placeholders (N/A, not_applicable), crop-cycle stages (harvest, sowing, storage) and stray non-season artefacts (lab, admin, math). The mapping folds rabi/boro/dry_season into winter and kharif/aman/flood into monsoon; aus, spring and pre_monsoon into summer. Values that were not seasons — cycle stages, multi-season combinations and artefacts — became all_year, which reads as "not season-specific". all_year is therefore a wide bucket: treat it as an absence of seasonal signal, not as a positive claim of year-round relevance.

---

How the dataset was built

Source material

Entries are grounded in publicly available Bangladeshi agricultural publications:

• BARI (Bangladesh Agricultural Research Institute) — crop cultivation guides, pest management manuals
• BRRI (Bangladesh Rice Research Institute) — rice variety catalogues and cultivation practice
• DAE (Department of Agricultural Extension) — extension field guides
• Banglapedia — the National Encyclopedia of Bangladesh, published by the Asiatic Society of Bangladesh; the largest single source, especially for soil science
• SRDI, BADC, BARC, DoF, DLS, BMD, BBS — soil resources, seed and input supply, research coordination, fisheries, livestock, meteorology and national statistics

Bengali is used throughout. English terms appear only where they are standard in Bangladeshi agricultural practice (NPK, pH, hybrid, dS/m).

Generation pipeline

Each source segment went through a structured pipeline. Questions were written at three difficulty tiers. For every question a reasoning trace was written — not a restatement of the answer, but the steps a knowledgeable agriculturist would actually follow. Answers were written in Bengali with an instructional tone, favouring practical recommendations, specific quantities and locally relevant examples. Every entry was then tagged with category, difficulty, season, topic, source and source type.

Disclosure: question generation, reasoning-trace construction and Bengali answer drafting were performed using Claude Sonnet 4.5 (Anthropic), with every entry grounded in the source material above. All entries went through human review before inclusion. The 2,094 entries labelled synthetic were generated from institutional knowledge spanning several references rather than from a single passage; this was used to fill coverage gaps in categories with fewer Bengali-language reference texts, such as agricultural economics and organic farming.

Entries are newly written rather than transcribed from the source documents.

Quality control

Three layers, each aimed at a different failure mode.

Layer 1 — generation-time review. Batches were spot-checked by a Bengali-speaking reviewer for untranslated English tokens, N/A placeholders, corrupted glyphs (Devanagari digits appearing in numbered lists), fragment answers that do not form a sentence, and factual drift from the source. Detected problems triggered regeneration; where a defect was systematic, the prompt or template was revised. An early soil-profile template that emitted N/A placeholders and untranslated English was caught this way, retired, and its entries regenerated.

Layer 2 — programmatic validator. A deterministic validator over the full corpus confirmed that all 5,529 entries parse as valid JSON, carry every required field, use a category from the eight-category taxonomy and a difficulty from the canonical tier set, and cite a non-empty source. Structural pass rate is 100%.

Layer 3 — human review. A category-stratified random sample of 200 entries was read in full — instruction, thinking and response — and scored by a Bengali-speaking agricultural reviewer on accuracy (binary), fluency (1–5) and readability (1–5).

Human validation results

Score	5	4	3	2	1	Mean
Fluency	144	28	18	6	4	4.51 / 5
Readability	148	24	14	13	1	4.53 / 5

Accuracy: 193 / 200 (96.5%). Equivalently, 3.5% of sampled entries needed correction or removal.

As reported in the paper, the 7 inaccurate entries fall into three reproducible defect types — fragment answers that do not form a complete sentence (3 entries, all in Soil & Nutrient Management), placeholder non-answers that restate the question (2 entries), and entries with garbled instructions where the question itself is meaningless (2 entries in the tech_* family) — and five of the eight categories scored 100% accuracy in the sample, with Soil & Nutrient Management weakest at 90.7%.

An earlier validation pass on a prior revision found a 10.5% inaccuracy rate, dominated by a single broken soil-profile template. That template was retired and its entries regenerated, which is what brought the rate down to 3.5%.

The sample and the reviewer's raw scores ship with the dataset (human_validation and human_validation_scores configs), so the score distributions above can be recomputed. Note that the sample is drawn from the training data — it is an audit sample, not a held-out evaluation split.

---

Findings: does it help?

We fine-tuned Qwen3-14B (4-bit) with LoRA on KrishiGyan and compared it against the base model. Full detail is in the paper; the headline is that every metric improved and seven of eight gains are statistically significant.

Setup. 5,528 complete examples were split into train (4,699), validation (276) and test (553) with a deterministic, category-stratified split. Metrics are reported on a fixed 100-example sample of the test split that preserves its category distribution. Targets concatenated the reasoning trace inside <think>…</think> followed by the answer; during scoring only the text after </think> was compared against the reference.

Training configuration

Parameter	Value
Base model	unsloth/Qwen3-14B-unsloth-bnb-4bit
Method	LoRA SFT (Unsloth), bf16 compute
GPU	NVIDIA H100 80 GB HBM3
Trainable parameters	128.45 M (0.8623%)
Max sequence length	2,048 tokens
LoRA rank / alpha / dropout	32 / 64 / 0.0
Epochs	2
Effective batch size	16
Learning rate	2 × 10⁻⁴, cosine schedule
Optimizer	AdamW 8-bit
Final training loss	0.2926

Base vs. fine-tuned

Confidence intervals are paired bootstrap 95% intervals over per-item differences (10,000 resamples); p is a two-sided Wilcoxon signed-rank test.

Metric	Base	Fine-tuned	Δ	95% CI
ROUGE-1	0.0763	0.1256	+0.0493	[.019, .083]
ROUGE-2	0.0405	0.0716	+0.0311	[.008, .057]
ROUGE-L	0.0763	0.1256	+0.0493	[.019, .083]
BLEU	0.0293	0.1111	+0.0818	[.051, .117]
chrF	0.2748	0.3470	+0.0722	[.036, .112]
METEOR	0.3801	0.4046	+0.0245	[−.017, .069] †
Semantic cosine	0.6983	0.7651	+0.0667	[.040, .095]
BERTScore F1	0.8560	0.8876	+0.0316	[.023, .041]

† METEOR is the exception — its interval spans zero (p = 0.62), so we make no claim of a reliable METEOR gain. All other metrics have p < 0.05.

Because agricultural advice can be correct while phrased differently from a reference, semantic cosine and BERTScore F1 are the metrics we treat as most informative. Both improve well clear of zero. The large BLEU jump mostly reflects a base model that was far from the reference surface form; the fine-tuned model adopts the concise, practical register of the KrishiGyan answers.

Gains by difficulty

Difficulty	n	ΔROUGE-L	ΔchrF	ΔCosine	ΔBERTScore F1
Easy	26	+0.0501	+0.1423	+0.0774	+0.0406
Medium	54	+0.0449	+0.0182	+0.0379	+0.0199
Hard	20	+0.0600	+0.1269	+0.1308	+0.0516

Hard questions benefit most. That is the pattern you would hope for from reasoning-augmented supervision: the gains concentrate exactly where multi-step domain reasoning is required, which is also where a wrong answer costs the most in practice.

Gains by category

Category	n	ΔROUGE-L	ΔchrF	ΔCosine	ΔBERTScore
Soil & Nutrient Mgmt.	29	+0.0552	+0.1621	+0.1329	+0.0518
Crop Prod. & Protection	20	+0.0892	+0.0886	+0.0440	+0.0411
Research & Extension	10	+0.1807	+0.0796	+0.0454	+0.0294
Technology & Innovation	13	+0.0000	+0.0291	+0.0739	+0.0177
Livestock & Fisheries	9	+0.0000	+0.0015	+0.0329	+0.0032
Organic & Sustainable	4	−0.1000	+0.0143	+0.0901	+0.0053
Water, Irrigation & Climate	9	+0.0153	+0.0060	−0.0128	+0.0290
Economics, Policy & Social	6	+0.0000	−0.0918	−0.0031	+0.0003

Several categories rest on as few as four evaluated items. Read this table as diagnostic, not definitive.

What changed qualitatively

Manual review of predictions shows a consistent shift. The base model produces plausible but generic Bengali that sometimes drifts into unsupported institutional detail. The fine-tuned model adopts the concise practical style of the references and uses Bangladesh-specific terminology correctly.

	Base model	Fine-tuned
"What is the salinity level of saline soil?"	Geographic background on coastal salt accumulation; cites "2 g/kg" — wrong unit and threshold	"Saline soil has salinity above 4 dS/m" — the reference threshold and unit
"Which rice varieties suit saline land?"	Ten invented variety names ("Bangla Saloyana", "Bangla Saloyana-2") that appear in no BRRI catalogue	"BRRI dhan47, BRRI dhan67, BRRI dhan73, BRRI dhan97, BINA dhan8 and 10 (tolerant up to 8–12 dS/m)"
"Which was established first, BARI or BRRI?"	"BRRI 1972, BARI 1976; difference 4 years" — both wrong	"BRRI 1970, BARI 1976; difference 6 years" — correct

Fine-tuning does not eliminate errors. Some outputs become too terse for economics and policy questions, and a few water and irrigation answers lose important nuance. Expert review remains necessary before any farmer-facing deployment.

---

Intended uses

Suitable for:

• Supervised fine-tuning of Bengali or multilingual LLMs for agricultural question answering
• Research on chain-of-thought supervision in low-resource languages
• Curriculum learning and per-tier analysis, using the difficulty labels
• Retrieval-augmented agricultural assistants for Bangladesh
• Evaluation and benchmarking of Bengali domain reasoning

Not suitable for:

• Direct farmer-facing advice without expert review in the loop
• Agricultural guidance outside Bangladesh, including West Bengal (see limitations)
• Any decision involving pesticide dosing, veterinary treatment or legal/policy compliance without verification against current official sources

---

Limitations

Geographic scope. Content is specific to Bangladesh. Soil classifications, crop varieties, institutional structures and climate conditions here do not transfer to West Bengal or other Bengali-speaking regions without adaptation.

Temporal currency. The dataset reflects knowledge current as of early 2026. Variety recommendations, pest prevalence and policy programmes shift over time.

Category balance. Crop production and soil management claim roughly 46% of entries; economics and organic agriculture about 7–8% each. This mirrors the available source material but is likely to produce weaker model performance in the underrepresented categories.

Synthetic content. 37.87% of entries are synthetically generated and grounded in institutional sources rather than extracted from a single passage. Synthetic text can carry patterns that differ from naturally occurring writing, and models trained on it may inherit them.

Residual errors. Human validation puts the residual error rate at 3.5%. The known defect types are listed above. Two entries with garbled instructions in the tech_* family are known to remain in the corpus.

Evaluation scale. The baseline metrics come from a 100-example sample of the 553-example test split, because the study ran on time-limited access to a single H100 and full-split inference with both a base and a fine-tuned 14B model across eight metrics was outside that budget. Confidence intervals and signed-rank tests are reported so the resulting uncertainty is visible rather than hidden, but 100 examples cannot establish how the gains generalise. Full-test-split evaluation is the first priority for follow-up work.

Metadata hygiene. 211 entries carry legacy difficulty aliases (basic, intermediate, advanced, high) rather than the three canonical tiers. These are retained deliberately: the paper reports the difficulty distribution with the aliases broken out as their own row, and maps them to core tiers at training time, so the released data matches the distribution reported there. The one-line mapping is given above. Separately, season was normalised from 110 raw values to four, and 70.79% of entries land in all_year — a bucket that absorbs both genuinely year-round entries and entries whose original annotation carried no usable seasonal signal. Do not read all_year as a strong claim.

Single-annotator validation. The 200-entry sample was reviewed by one Bengali-speaking agricultural reviewer, so no inter-annotator agreement can be reported.

---

Safety

Agricultural advice carries real risk when it is wrong. This dataset contains pesticide dosing, fertiliser rates, disease treatment guidance and policy claims. Any deployed system built on it should include retrieval from verifiable current sources, explicit uncertainty warnings, and escalation to qualified agricultural experts for high-risk recommendations. Do not treat model output trained on this data as a substitute for extension services.

---

Licensing and attribution

The dataset

KrishiGyan is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).

You are free to share (copy and redistribute in any medium or format) and adapt (remix, transform and build upon the material), for any purpose, including commercially — provided you meet the attribution terms:

• Attribution. Give appropriate credit, provide a link to the licence, and indicate whether changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensors endorse you or your use.
• No additional restrictions. You may not apply legal terms or technological measures that legally restrict others from doing anything the licence permits.

A minimal attribution line:

KrishiGyan dataset by Aparna Roy and Shudipto Trafder, licensed under CC BY 4.0. Available at https://huggingface.co/datasets/AROY76/KrishiGyan

If you modify the data, say so — filtered subsets, normalised metadata and translated variants all count as changes under the licence.

Upstream sources

The KrishiGyan records — questions, reasoning traces and answers — are original text written for this dataset and are what the CC BY 4.0 grant covers. The underlying agronomic facts derive from publications by the institutions listed below, which retain all rights in their own original documents. Entries were written rather than transcribed, but if you intend to redistribute or build products on the source publications themselves, consult the original publishers directly.

Source	Entries referencing it
DAE — Department of Agricultural Extension	1,464
Banglapedia — Asiatic Society of Bangladesh	1,177
BARI — Bangladesh Agricultural Research Institute	1,168
BRRI — Bangladesh Rice Research Institute	1,092
BADC — Bangladesh Agricultural Development Corporation	562
SRDI — Soil Resource Development Institute	518
DoF — Department of Fisheries	368
BARC — Bangladesh Agricultural Research Council	358
BMD / BMDA — meteorology and barind development	311
BBS — Bangladesh Bureau of Statistics	223
DLS — Department of Livestock Services	216
BFRI — Bangladesh Fisheries / Forest Research Institute	167

(Counts are mentions across the 441 distinct source strings; an entry citing several institutions is counted under each.)

A small number of entries (15) reference Wikipedia, whose content is CC BY-SA. As above, those entries are newly written rather than copied.

Generated content

Portions of this dataset were produced with Claude Sonnet 4.5 (Anthropic) under human review, as described in Generation pipeline. If your use case requires wholly human-authored training data, this dataset is not appropriate for it.

---

Citation

If you use KrishiGyan, please cite the paper. It is accepted and forthcoming at PEEIACON 2026; this entry will be updated with the final page numbers and DOI once the proceedings are published.

@inproceedings{roy2026krishigyan,
  title     = {KrishiGyan: A Chain-of-Thought Bengali {Q\&A} Dataset for
               Bangladesh Agriculture with a {Qwen3-14B} {LoRA} Baseline},
  author    = {Roy, Aparna and Trafder, Shudipto},
  booktitle = {2026 IEEE International Conference on Power, Electrical,
               Electronics and Industrial Applications (PEEIACON)},
  address   = {Dhaka, Bangladesh},
  publisher = {IEEE},
  year      = {2026},
  note      = {To appear}
}

And, if you cite the dataset artefact directly:

@misc{krishigyan2026data,
  title        = {KrishiGyan: A Chain-of-Thought Bengali Agricultural
                  {Q\&A} Dataset for Bangladesh},
  author       = {Roy, Aparna and Trafder, Shudipto},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/AROY76/KrishiGyan}},
  note         = {Licensed under CC BY 4.0}
}

---

Acknowledgements

To BARI, BRRI, DAE, SRDI, BADC, BARC, DoF, DLS, BMD and BBS, whose published research and extension material made this dataset possible, and to the Asiatic Society of Bangladesh for Banglapedia.

---

Versions

v1.0 — the release that accompanies the PEEIACON 2026 paper. Its statistics correspond to the tables reported there, so figures quoted in the paper and figures computed from these files line up.

Later revisions will refine metadata labelling — topic and category assignment in particular — and will restate any statistic that changes as a result. Where a future version and the paper disagree, the version notes for that release are the authority on what changed and why; the paper remains the record of the dataset as evaluated.

---

Roadmap

• Metadata labelling pass: content-derived category and topic assignment
• Full 553-example test-split evaluation with confidence intervals
• Ablations: with and without reasoning traces; curriculum versus random training order
• Additional model families — Bangla-specialised models, LLaMA, Mistral, Gemma
• Expansion toward 20,000 entries, weighted to the underrepresented categories
• Expert assessment by agricultural scientists and extension workers in Bangladesh
• An English parallel corpus for cross-lingual transfer research
• A field pilot with farmers in rural Bangladesh

Corrections and issue reports are welcome through the Community tab.