Marconcini commited on
Commit
203f616
·
verified ·
1 Parent(s): a2fca43

Model card: demo link and use in the cited assistant

Browse files
Files changed (1) hide show
  1. README.md +18 -0
README.md CHANGED
@@ -66,6 +66,24 @@ Files:
66
  In the browser (ONNX Runtime Web, image resizing done by the canvas instead of PIL), the demo
67
  agrees with the Python int8 predictions on 96.1% of PlantDoc test photos.
68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69
  ## Intended use and limitations
70
 
71
  - **Intended use:** demonstration and research on robustness of plant-disease classifiers.
 
66
  In the browser (ONNX Runtime Web, image resizing done by the canvas instead of PIL), the demo
67
  agrees with the Python int8 predictions on 96.1% of PlantDoc test photos.
68
 
69
+ ## In the LeafScan app: diagnosis, then a cited assistant
70
+
71
+ 👉 Try the model in your browser: [Marconcini/leafscan](https://huggingface.co/spaces/Marconcini/leafscan)
72
+ (int8 ONNX, nothing uploaded).
73
+
74
+ The project's local app goes one step further: after the diagnosis, a retrieval-augmented
75
+ assistant answers questions from 28 Wikipedia articles with a local LLM, citing its sources
76
+ (`LEAFSCAN_ASSISTANT=1 python -m leafscan.app`). How this model's output is used there was
77
+ tested on 500 (question, field photo) pairs with its real predictions on PlantDoc test:
78
+
79
+ - Restricting the search to the predicted disease **costs 16 points of recall@5**: a right
80
+ diagnosis adds 8 points, a wrong one (28% of field photos) removes the answer.
81
+ - Restricting to the top-2 when the model is unsure recovers less than half of the loss.
82
+
83
+ So the diagnosis does **not** filter the search: it suggests questions that name the disease
84
+ (both candidates when the model says "Not sure"), which the user can see and correct.
85
+ Details: [research log](https://github.com/marconcini/leafscan/blob/main/docs/research-log.md).
86
+
87
  ## Intended use and limitations
88
 
89
  - **Intended use:** demonstration and research on robustness of plant-disease classifiers.