README.md

---
language:
- en
- es
task_categories:
- text-generation
- question-answering
tags:
- agriculture
- climate
- crop-calendar
- evidence-grounding
- bilingual
- autoscientist
- adaption
license:
- cc-by-4.0
- cc0-1.0
pretty_name: AgroVeritas Evidence QA — Adapted Training Dataset
size_categories:
- 10K<n<100K
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: viewer/agroveritas_train.jsonl
---

# AgroVeritas Evidence QA

> Agricultural intelligence you can audit — in English and Spanish.

AgroVeritas is a bilingual, evidence-bounded agricultural instruction dataset built for regional crop-calendar and historical climate reasoning. It teaches models to answer a practical question, cite the evidence used, show the reasoning path, state limitations, and recommend local verification instead of presenting historical data as live field conditions.

## Dataset Viewer and canonical export

The default Hugging Face Dataset Viewer reads `viewer/agroveritas_train.jsonl`. This viewer-safe mirror preserves all 28,220 rows, their order, and every non-null value while representing optional null metadata as empty strings so every column has a stable string schema.

The canonical byte-for-byte Adaption export remains available as `agroveritas_adapted_training_711828b2.jsonl`. Its published SHA-256 remains `c0f6042e548a6daf62cb7f988ad8b172014548fad5575abd63eb123e45dd402d`.

This repository is the public release of the **adapted dataset used by AutoScientist** for the winning AgroVeritas model run. The release is intentionally paired with its source provenance, transformation record, quality evidence, and exact training configuration.

## Headline results

| Signal | Before | After | Change |
|---|---:|---:|---:|
| Final Adaption quality score | 9.0 / 10 | **10.0 / 10** | **+11.1% relative** |
| Final Adaption grade | A | **A** | maximum score reached |
| Final Adaption percentile | 43.9 | **57.7** | **+13.8 points** |
| Agriculture held-out win rate | 23% base | **78% adapted** | **+55 points; +239% relative lift** |

The broader project iteration journey began with a 2,960-row prototype graded **C, 5.0/10, 6.9th percentile** and reached the final **A, 10.0/10, 57.7th percentile** corpus. That is a **+100% quality-score gain** and a **+50.8 percentile-point rise across successive dataset versions**. It is not presented as a single-run before/after measurement; the table above contains the final same-run comparison.

## What is in the release

- **28,220 exact exported training rows** from Adaption dataset `711828b2-e610-4ce4-a4ef-3ef4244f7534`.
- **17,094 agriculture-core rows** plus **11,126 AutoScientist expansion rows**—a **39.4% general-purpose diversity buffer** designed to reduce catastrophic narrowing and preserve instruction-following breadth.
- Original prompt, original completion, and enhanced completion fields preserved by the platform export.
- English and Latin American Spanish instructions.
- Regional crop-calendar, date-membership, climate-profile, variable, comparison, extreme, and guarded advisory tasks.
- Row-level source IDs, source URLs, language, region, crop, license, verification state, and evidence policy where present in the source record.

The source-generation library contained 33,250 strictly validated rows and 6,383 unique fact groups. AutoScientist used the 28,220-row exported training artifact for supervised fine-tuning. Platform-added augmentation remains identifiable rather than being misrepresented as FAO or NASA data: expansion rows have platform-generated `enhanced_prompt` and `enhanced_completion` values and do not carry AgroVeritas source metadata.

### Exact training export composition

| Partition | Rows | Share |
|---|---:|---:|
| Agriculture core | 17,094 | 60.6% |
| AutoScientist general-purpose expansion | 11,126 | 39.4% |
| **Total exact training export** | **28,220** | **100%** |

All 28,220 rows contain an `enhanced_completion`. The 17,094 agriculture-core rows preserve their original prompt, original completion, domain metadata, and source lineage. The 11,126 expansion rows contain an `enhanced_prompt` and `enhanced_completion`; their domain-specific metadata fields are intentionally null.

## The five-part evidence contract

Every domain answer is trained toward a consistent contract:

1. **Answer in scope** — respond only to what the cited record supports.
2. **Evidence** — identify the source and the exact crop-calendar or climate value used.
3. **Reasoning** — connect the evidence to the conclusion without hidden assumptions.
4. **Limitations** — distinguish historical climatology and published calendars from live weather or field observations.
5. **Local confirmation** — recommend verification with local extension services or current authoritative data before operational decisions.

## Source corpus profile

| Dimension | Count |
|---|---:|
| Source rows | 33,250 |
| Unique prompts | 33,250 |
| Unique completions | 33,235 |
| Unique fact groups | 6,383 |
| English (`en-US`) | 17,810 |
| Spanish (`es-419`) | 15,440 |

### Task distribution

| Task | Rows |
|---|---:|
| Regional crop calendar | 12,453 |
| Calendar date membership | 12,453 |
| Regional climate profile | 2,760 |
| Regional climate variable | 2,760 |
| Regional climate advisory with guardrails | 2,072 |
| Regional climate comparison | 564 |
| Regional climate extreme | 188 |

## Provenance and licensing

The agricultural records are derived from openly licensed, attributed sources:

- **FAO Crop Calendar** records — source attribution preserved; CC BY 4.0.
- **NASA POWER** historical climatology — public-domain/CC0-compatible source terms.
- **GeoNames** geographic identifiers — CC BY 4.0.

Generated instructional framing and transformations are released under CC BY 4.0. Each underlying record retains its own source and license metadata. This release is not endorsed by FAO, NASA, GeoNames, Meta, Hugging Face, Kaggle, or Adaption.

## Quality and leakage controls

- Strict validation status: `PASS_STRICT_FROZEN_UPLOAD_READY`.
- Training rows were grouped by `fact_group`, not split as isolated bilingual twins.
- Fact-group overlap with the private holdout: **0**.
- Fact-group overlap with the private evaluation set: **0**.
- Exact prompt overlap with the frozen holdout: **0**.
- Source output SHA-256: `988bd94db112f6b2cf5d9126385cdf760bf51ff573496583e3704b1397d46d7a`.
- Exact adapted training export SHA-256: `c0f6042e548a6daf62cb7f988ad8b172014548fad5575abd63eb123e45dd402d`.

The private holdout and private evaluation rows are **not published and were not used for training**.

## Intended uses

- Bilingual agricultural question answering.
- Regional crop-calendar interpretation.
- Historical climate comparison and climate-variable explanation.
- Evidence-grounded agricultural assistants and research prototypes.
- Training and evaluating calibrated abstention when the source does not support a live or field-specific claim.

## Limitations and safety

This dataset does not provide live weather, field scouting, soil tests, pest diagnosis, pesticide prescriptions, financial advice, or guarantees of yield. Historical climatology and published crop calendars can be outdated or locally incomplete. Outputs should support—not replace—local agronomists, extension services, current forecasts, label instructions, and applicable law.

## Reproducibility

See `dataset_metrics.json`, `LICENSES.md`, and the [companion AgroVeritas model](https://huggingface.co/MarianaCodebase/AgroVeritas-Scout-17B). The exact AutoScientist recipe is included with the model release. Screenshots of the 10/10 dataset evaluation and 78% Agriculture win rate are included in the project submission evidence bundle.

## Citation

```bibtex
@dataset{agroveritas2026,
  title        = {AgroVeritas Evidence QA: Bilingual Evidence-Bounded Agriculture Instructions},
  author       = {Sinisterra, Mariana},
  year         = {2026},
  note         = {Created with Adaptive Data and AutoScientist by Adaption}
}
```