| --- |
| title: AgroVeritas Evidence Assistant |
| emoji: 🌾 |
| colorFrom: green |
| colorTo: blue |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: cc-by-4.0 |
| short_description: Auditable bilingual agriculture with FAO and NASA evidence |
| models: |
| - MarianaCodebase/AgroVeritas-Scout-17B |
| datasets: |
| - MarianaCodebase/AgroVeritas-Evidence-QA-Adapted |
| tags: |
| - agriculture |
| - bilingual |
| - evidence-grounding |
| - autoscientist |
| --- |
| |
| # AgroVeritas Evidence Assistant |
|
|
| A public bilingual interface for AgroVeritas, the AutoScientist Agriculture project by Mariana Sinisterra. |
|
|
| The Space demonstrates the model's five-part evidence contract using verified public scenarios: |
|
|
| 1. agricultural finding; |
| 2. supporting record; |
| 3. reproducible verification logic; |
| 4. evidence limitations; |
| 5. local extension review. |
|
|
| ## Verified project results |
|
|
| - **78%** Agriculture held-out win rate versus **23%** for the frozen baseline. |
| - **A / 10.0** final dataset quality. |
| - **28,220** exact adapted training rows. |
| - English and Latin American Spanish. |
|
|
| This static Space is a curated evidence-retrieval preview. It does not claim to serve the 109B-parameter Llama 4 Scout base checkpoint. The exact PEFT adapter, training recipe, evaluation record, and dataset are linked from the interface. |
|
|
| Built with Adaptive Data and AutoScientist by Adaption. |
|
|
