v2: quality-gate + VLM content filter fix (6,066 records, 7 categories, replaces prior junk-contaminated release)

README.md

@@ -28,11 +28,22 @@ A sample of farmer-submitted photographs from FarmerChat, an agricultural adviso
 used by smallholder farmers in India, Ethiopia, Kenya and Nigeria. Published by
 [Digital Green](https://www.digitalgreen.org).
 
-This release contains 8,000 records (7,906 distinct photographs; some photographs belong to
-more than one category, see below) drawn from 8 categories representing different outcomes of
-an automated crop/livestock diagnosis pipeline. Roughly 1,000 records per category, sampled
+This release contains 6,066 records (5,930 distinct photographs; some
+photographs belong to more than one category, see below) drawn from 7 categories
+representing different outcomes of an automated crop/livestock diagnosis pipeline, sampled
 across country, month, crop and diagnosis where each applies.
 
+Submissions span 2025-08 to 2026-07.
+
+## Changes in this version
+
+This version replaces an earlier release that had two problems: some images had never
+actually cleared the service's quality gate (most notably in the livestock category), and
+a small number of off-topic images (screenshots, AI-generated pictures) had slipped
+through because the quality gate only checks blur/framing, not content. Both are fixed
+here — see "How these images were selected" below — which is why the record count and
+category sizes differ from the previous version.
+
 ## Loading
 
 ```python
@@ -44,22 +55,41 @@ ds = load_dataset("DigiGreen/farmerchat-image-samples", split="train")
 Or load `metadata.csv` directly and read images from the `images/` folder using its
 `file_name` column.
 
+## How these images were selected
+
+Every photograph in this release cleared two independent screens.
+
+1. **The service's own image-quality gate.** Submissions that the gate rejected (blur,
+   lighting, framing, no plant detected) are not included. This restricts the release to
+   submissions from the period the gate has been running.
+2. **An automated content check.** Each remaining image was reviewed by a vision-language
+   model and kept only if it is genuinely agricultural, is a plausible diagnostic
+   submission rather than an accidental or off-topic one, and is an original photograph
+   rather than a screenshot, a stock or template graphic, or an AI-generated image.
+
+Farmers do send off-topic images to the service — screenshots, photos of documents, pets,
+food, AI-generated pictures. Those are a real and interesting property of the traffic, but
+they are filtered out here so this dataset is what it claims to be: real photographs of
+crops and livestock.
+
 ## Categories
 
-| Category | What it represents |
-|---|---|
-| `diagnosis_returned` | A crop was identified and a specific disease/pest diagnosis was returned |
-| `crop_identified_diagnosis_unresolved` | A crop was identified but the health outcome was not determined |
-| `crop_unresolved` | No crop could be identified in the photograph |
-| `rejected_quality` | The photograph was rejected by an automated image-quality check (blur, lighting, framing, etc.) before diagnosis was attempted |
-| `healthy_crop` | A crop was identified and assessed as healthy |
-| `low_confidence_diagnosis` | A diagnosis was returned, but with lower model confidence |
-| `livestock` | The photograph was flagged as livestock-related rather than crop-related |
-| `multimodal_organic_text` | The photograph was accompanied by farmer-written text (not a repeated quick-reply template) |
+| Category | Records | What it represents |
+|---|---:|---|
+| `diagnosis_returned` | 991 | A crop was identified and a specific disease/pest diagnosis was returned |
+| `crop_identified_diagnosis_unresolved` | 991 | A crop was identified but the health outcome was not determined |
+| `crop_unresolved` | 983 | No crop could be identified in the photograph |
+| `healthy_crop` | 987 | A crop was identified and assessed as healthy |
+| `low_confidence_diagnosis` | 995 | A diagnosis was returned, but with lower model confidence |
+| `livestock` | 120 | The submission was flagged as livestock-related rather than crop-related |
+| `multimodal_organic_text` | 999 | The photograph was accompanied by farmer-written text (not a repeated quick-reply template) |
 
 An image can belong to more than one category (for example, a livestock photo with a
-low-confidence diagnosis). Each membership is its own row in `metadata.csv`, sharing the same
-`file_name`.
+low-confidence diagnosis). Each membership is its own row in `metadata.csv`, sharing the
+same `file_name`.
+
+Category sizes are uneven. `livestock` is much smaller than the others because most
+livestock-related submissions do not clear the image-quality gate.
 
 ## Columns
 
@@ -67,7 +97,7 @@ low-confidence diagnosis). Each membership is its own row in `metadata.csv`, sha
 |---|---|
 | `file_name` | Relative path to the image file |
 | `image_id` | Stable per-image identifier |
-| `category` | One of the 8 categories above |
+| `category` | One of the categories above |
 | `country` | Country the submission came from |
 | `month` | Year-month the photo was submitted (YYYY-MM) |
 | `crop` | Crop identified in the photo, where resolved. Normalized to a standard set of names (synonyms merged, casing standardized) |
@@ -88,13 +118,17 @@ automated language-model pass; any personal names, phone numbers, email addresse
 addresses found were replaced with `[REDACTED]`.
 
 Automated screening reduces but does not eliminate the chance of residual personal content.
-This dataset should not be treated as guaranteed free of it.
+This dataset should not be treated as guaranteed free of it. In particular, some photographs
+carry a camera watermark or timestamp burnt into the image by the phone that took them.
 
 ## Limitations
 
 Diagnoses and crop/health labels are outputs of an automated pipeline, not verified expert
 labels. They reflect what the system determined, not necessarily agronomic ground truth.
 
+The content check in step 2 above is itself an automated model judgement, so it will have
+made some mistakes in both directions.
+
 ## License
 
 CC BY 4.0.