Link public fine-tuning code and chart data

EVALUATION.md

@@ -4,6 +4,8 @@ This page reports the **selected Stage-B step-1,000 checkpoint**, SHA-256 `ef807
 
 ![Aural One audio evaluation chart: macro-F1 is shown separately from question accuracy](assets/audio-evaluation.png)
 
+The chart is generated from [`audio_eval_chart_data.json`](assets/audio_eval_chart_data.json) with `python scripts/render_audio_eval.py` (requires Matplotlib). Emotion macro-F1 and question accuracy use separate panels because their denominators and meanings differ.
+
 ## Emotion and typed decisions
 
 | Fixed evaluation | Frozen step-420 reference | Aural One preview | Scope |

TRAINING.md

@@ -17,3 +17,5 @@ The training data sources were:
 The audio files and speaker-level manifests are **not included** in either public repository. Evaluation datasets have their own terms. For source-disjoint speaker protocols, sample counts, and external checks, see [EVALUATION.md](EVALUATION.md).
 
 The published delta is SHA-256 `ef80763236b2467a886d52fba51769de4dcfbdce909dd320803b6d2d2d41db96` for the acoustic weights and `e2b53154b40cd67faf3c9a57226f09c187b060569a7894ee2b3630a4e88937c3` for the adapter. `release.json` pins the base, file hashes, update count, and seed. The base weights are downloaded directly from Google's repository.
+
+The reusable training scripts, manifest schema, and checkpoint/evaluation commands are in [FINE_TUNING.md](https://github.com/Parassharmaa/aural-one/blob/main/docs/FINE_TUNING.md). The public schedule builder regenerated the selected 1,000-update schedule's hash exactly from the original local manifests.

assets/audio_eval_chart_data.json

@@ -0,0 +1,14 @@
+{
+  "emotion_macro_f1": [
+    {"label": "SUBESCO development", "scope": "700 clips · 2 unseen speakers", "reference": 0.134, "aural_one": 0.295},
+    {"label": "CREMA-D development", "scope": "445 clips · unseen actors", "reference": 0.361, "aural_one": 0.391},
+    {"label": "SUBESCO reserved test", "scope": "1,012 consensus clips · 4 unseen speakers", "reference": null, "aural_one": 0.344},
+    {"label": "German EmoDB", "scope": "195 clips · previously opened", "reference": 0.078, "aural_one": 0.102},
+    {"label": "Italian Emozionalmente", "scope": "262 unanimous clips · development actors", "reference": 0.232, "aural_one": 0.203}
+  ],
+  "audio_question_accuracy": [
+    {"label": "Draft 60s speech questions", "scope": "6 recordings · development", "correct": 14, "total": 18},
+    {"label": "Initial sound screen", "scope": "24 clips · mixed recording sources", "correct": 14, "total": 24},
+    {"label": "Natural 45–60s speech", "scope": "39 distinct held-out recordings", "correct": 18, "total": 39}
+  ]
+}