Add clusters.json

manifests/clusters.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 6,
-  "description": "Atomic architectural-lineage clusters for the 190-case ONNX benchmark. Each case belongs to exactly one cluster; each cluster's members share a sub-family; each sub-family maps to exactly one family. Clusters are the atomic unit for train/val/test splits (no cluster straddles two splits). Schema v6: cluster boundaries tightened to enforce the bigram-set equivalence rule (see docs/taxonomy-rationale.md \u00a73) \u2014 19 clusters split/unmerged, 43 new clusters created (total 156). Un-done schema-v5 \u00a7B merges that fail the bigram Jaccard < 0.40 test: bit \u2194 ResNet_lineage (J=0.125); openai_gpt \u2194 GPT_lineage (J=0.42). See docs/cluster-decision-log.md for the full per-cluster record.",
+  "description": "Atomic architectural-lineage clusters for the 190-case ONNX dataset. Each case belongs to exactly one cluster; each cluster's members share a sub-family; each sub-family maps to exactly one family. Clusters are the atomic unit for train/val/test splits (no cluster straddles two splits). Schema v6: cluster boundaries tightened to enforce the bigram-set equivalence rule (see docs/taxonomy-rationale.md \u00a73) \u2014 19 clusters split/unmerged, 43 new clusters created (total 156). Un-done schema-v5 \u00a7B merges that fail the bigram Jaccard < 0.40 test: bit \u2194 ResNet_lineage (J=0.125); openai_gpt \u2194 GPT_lineage (J=0.42). See docs/cluster-decision-log.md for the full per-cluster record.",
   "subfamily_to_family": {
     "audio_encoder": "attention_encoder",
     "bart_family": "attention_encoder_decoder",