Upload baselines/asr-africa__w2v2-bert-wolof-kallaama-42-hours__bocalantics-wolof-eval.json with huggingface_hub

baselines/asr-africa__w2v2-bert-wolof-kallaama-42-hours__bocalantics-wolof-eval.json

@@ -0,0 +1,43 @@
+{
+  "eval_set": "MOH749/bocalantics-wolof-eval",
+  "interpretation": "Per-source matters more than the headline here. This model was trained on Kallaama, and Kallaama rows are part of this eval set - a much lower WER on those than on the others is memorisation of its own training distribution, not generalisation to Wolof. Compare the sources before quoting the total.",
+  "model": "asr-africa/w2v2-bert-wolof-kallaama-42-hours",
+  "overall": {
+    "cer": 0.2663,
+    "wer": 0.513
+  },
+  "per_source": {
+    "MOH749/kallaama": {
+      "cer": 0.2213,
+      "in_model_training_data": true,
+      "n": 698,
+      "wer": 0.3518
+    },
+    "galsenai/wolof-audio-data": {
+      "cer": 0.1779,
+      "in_model_training_data": false,
+      "n": 1751,
+      "wer": 0.4018
+    },
+    "serge-wilson/wolof-french-asr": {
+      "cer": 0.4622,
+      "in_model_training_data": false,
+      "n": 1213,
+      "wer": 0.9116
+    },
+    "soynade-research/Wolof-ASR-Data": {
+      "cer": 0.2094,
+      "in_model_training_data": false,
+      "n": 982,
+      "wer": 0.4816
+    },
+    "vonewman/wolof-audio-data": {
+      "cer": 0.1743,
+      "in_model_training_data": false,
+      "n": 126,
+      "wer": 0.3851
+    }
+  },
+  "rows": 4770,
+  "schema_version": "yaatal.baseline-eval.v1"
+}