Honest labels: eval is agent-labelled, not human-reviewed

README.md

@@ -20,30 +20,13 @@ source_datasets:
 pretty_name: Southborough Death Cards (gold / silver / pool)
 configs:
 - config_name: eval
-  data_files:
-  - split: train
-    path: eval/train-*
-- config_name: gold
-  data_files: gold/*
-- config_name: pool
-  data_files: pool/*
+  data_files: eval/*
 - config_name: silver
   data_files: silver/*
-dataset_info:
-  config_name: eval
-  features:
-  - name: image_id
-    dtype: string
-  - name: image
-    dtype: image
-  - name: extraction
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 1930935
-    num_examples: 12
-  download_size: 1931399
-  dataset_size: 1930935
+- config_name: pool
+  data_files: pool/*
+annotations_creators:
+- machine-generated
 ---
 
 # Southborough death cards — structured extraction
@@ -56,12 +39,16 @@ prepared for schema-conditioned extraction. Each card records a death: `name`, `
 
 | config | cards | labels |
 |---|---|---|
-| `gold` | 12 | **hand-reviewed** ground truth (eval) — AI-drafted, human-reviewed, uncertain readings flagged |
-| `silver` | 127 | **machine-generated** labels (Qwen3-VL-8B from the image) — SFT data, *not* hand-verified |
+| `eval` | 12 | **AI-labelled, not human-verified** — a Claude agent read each card image carefully (one at a time, flagging uncertain readings). Held out for evaluation. *Higher-quality model labels, but not gold.* |
+| `silver` | 127 | **machine-generated** labels (Qwen3-VL-8B, bulk from the image) — SFT data, *not* hand-verified |
 | `pool` | 160 | **unlabelled** image pool — re-label with your own model/schema |
 
-Training (`silver`) and evaluation (`gold`) data for the [index-card-extractor-4b](#) research model;
-`pool` is the unlabelled source for `silver`. ⚠️ `silver` labels carry model noise on free-text fields.
+Training (`silver`) and evaluation (`eval`) data for the [index-card-extractor-4b](#) research model;
+`pool` is the unlabelled source for `silver`.
+
+> ⚠️ **No human-verified ground truth here.** Both `eval` and `silver` are model-generated (the `eval`
+> set more carefully, by a stronger model). Evaluation numbers against `eval` are *relative to agent
+> labels*, not gold. Free-text fields (names, places) carry model noise.
 
 ## Columns
 `gold`/`silver`: `image_id` · `image` · `extraction` (JSON). `pool`: `image_id` · `image`.