Card: scrub stale cot_probe_hq entry; composition table with renamed groups

README.md

@@ -123,38 +123,6 @@ dataset_info:
     num_examples: 100
   download_size: 180567
   dataset_size: 269397
-- config_name: cot_probe_hq
-  features:
-  - name: id
-    dtype: string
-  - name: set
-    dtype: string
-  - name: title
-    dtype: string
-  - name: text
-    dtype: string
-  - name: origin
-    dtype: string
-  - name: source
-    dtype: string
-  - name: model
-    dtype: string
-  - name: env
-    dtype: string
-  - name: category
-    dtype: string
-  - name: label
-    dtype: string
-  - name: context
-    dtype: string
-  - name: axis
-    dtype: string
-  splits:
-  - name: eval
-    num_bytes: 71304
-    num_examples: 57
-  download_size: 38004
-  dataset_size: 71304
 - config_name: frontier_metagame_rollouts
   features:
   - name: id
@@ -321,10 +289,6 @@ configs:
   data_files:
   - split: eval
     path: claude_human_forget/eval-*
-- config_name: cot_probe_hq
-  data_files:
-  - split: eval
-    path: cot_probe_hq/eval-*
 - config_name: frontier_metagame_rollouts
   data_files:
   - split: eval
@@ -349,6 +313,32 @@ configs:
 
 # camgeodesic/metagaming-labeling
 
+The combined **metagaming token-labeling dataset**: documents and chat
+transcripts labeled (or to be labeled) for metagaming / evaluation-awareness
+content, used to train and validate token-level probes. Built by
+[`dataset-builder`](https://github.com/GeodesicResearch/dataset-builder)
+(`configs/metagaming_labeling/`, one folder per group).
+
+## Composition (~10.3M whitespace tokens, 7,568 docs)
+
+| Config | Docs | ~Tokens | What it is |
+|---|---|---|---|
+| `claude_AI_forget` | 132 | 84k | Claude-authored synthetic metagaming docs (forget) |
+| `claude_general_retain` | 100 | 39k | far-negative controls (retain) |
+| `claude_AI_mix` | 100 | 37k | ML-adjacent distractors |
+| `claude_human_forget` | 100 | 39k | human-domain oversight & evaluation docs (forget) |
+| `westover_metagame_split` | 193 | 354k | organic discourse, metagaming-relevant |
+| `westover_general_split` | 3,740 | 6.24M | organic discourse, background |
+| `kimi_metagame_rollouts` | 2,240 | 2.82M | Kimi-K3 chat transcripts, 13 arms, CoT inline (`User:`/`Assistant:` format) |
+| `frontier_metagame_rollouts` | 57 | 8k | REAL frontier-model metagame CoTs (o3 / o4-mini: Apollo arXiv 2509.15541 App N + OpenAI SA transcript) — validation-grade |
+| `needham_eval_awareness` | 906 | 708k | eval-awareness document set |
+| `annotations` | 117 | — | human forget-token spans (whitespace-token indices over sibling groups) |
+
+Token counts are whitespace (`\S+`) tokens over each group's `text` column —
+the same tokenization the `annotations` spans index. This section is
+hand-maintained; per-config schemas below are auto-generated and refresh on
+every push.
+
 Auto-generated by [`dataset-builder`](https://github.com/GeodesicResearch/dataset-builder).
 Each config below is a separate dataset produced from a versioned YAML build
 config. Load with: