Card

README.md

@@ -0,0 +1,32 @@
+---
+license: cc-by-4.0
+pretty_name: Hebrish Model Iterations
+tags: [asr, model-training, changelog]
+configs:
+- config_name: default
+  data_files: iterations.jsonl
+---
+
+# Hebrish Model Iterations
+
+One row per released iteration of every Hebrish speech model (Sabich, Shakshuka, Jachnun and variants), with the data that went into it. Newest last.
+Updated after every release.
+
+| Column | Meaning |
+|---|---|
+| `model`, `version`, `label`, `run` | Model, its version, the release label and the training run id |
+| `date` | Release (or training) date |
+| `base_model` | The checkpoint family the line was first fine-tuned from |
+| `method` | `incremental` (continued from the previous weights) or `from stock` |
+| `weights_from`, `previous_version`, `previous_run` | The run the weights came from, and the iteration the delta is measured against |
+| `blend` | Data blend: `all` (everything), `cs` (code-switching mix), `general` (everything except code-switching) |
+| `dataset_revision` | Commit of the recordings dataset the run trained on |
+| `pool_takes`, `pool_hours` | Takes and hours of the blend's training pool at that revision |
+| `delta_takes`, `delta_minutes`, `delta_hours` | Takes in the pool that were not in the previous iteration's pool |
+| `growth` | Pool minutes relative to the previous iteration's pool |
+| `new_terms` | Target terms new to the pool |
+| `epochs`, `lr`, `train_minutes`, `best_val_wer`, `gpu` | Training recipe and outcome (validation WER is per run and not comparable across runs) |
+| `hf_repo` | Where the weights are published |
+
+Pool figures apply today's blend rules to each old revision, so for older runs they are close to, not
+identical with, what the run trained on. A from-stock run with no earlier iteration counts its whole pool as delta.