Link the card to the GitHub repository and its siblings

README.md

@@ -1,14 +1,69 @@
 ---
-license: cc-by-nc-4.0
-language: [ln, sn]
-tags: [automatic-speech-recognition, waxal, zindi]
+license: apache-2.0
+language:
+- ln
+base_model: openai/whisper-large-v3
+library_name: transformers
+pipeline_tag: automatic-speech-recognition
+tags:
+- automatic-speech-recognition
+- waxal
+- zindi
+- african-languages
 ---
 
-# waxal-qlora-largev3-lin
+# Lingala ASR — whisper-large-v3 + QLoRA adapter
 
-lin QLoRA adapter over whisper-large-v3 (PEFT, load onto the base model)
+Component of the **4th-place solution** to the [Google WAXAL ASR Challenge](https://github.com/yehoshua0/waxal-asr-phase2) (Zindi, phase 2):
+892 unseen clips, two African languages, **no language metadata**, scored `1 - (WER + CER) / 2` on
+raw text. Private leaderboard **0.771848284**.
 
-Trained for the Google WAXAL ASR Challenge (Zindi), phase 2. Full method:
-`docs/METHOD.md` in the solution bundle. This card is a stub written at upload time and is
-replaced by the documented card in the same commit series.
+> **Code, full method and one-command verification: [yehoshua0/waxal-asr-phase2](https://github.com/yehoshua0/waxal-asr-phase2)**
+> The repository reproduces the submitted CSV byte for byte on a laptop in about a minute, and
+> re-decodes every input from the audio on rented GPUs in about four hours.
 
+## Role in the system
+
+**Witness (voter).** Load onto the base model with PEFT and keep the regime it was selected under: 4-bit NF4 base, adapter NOT merged, bf16 autocast. Merging into a full-precision base changes the arithmetic it was measured in.
+
+## What it measured
+
+Drives the final arbitration stage (`qlora_voter`, +0.000216). As a solo Lingala half: 0.742006.
+
+## Usage
+
+```python
+from peft import PeftModel
+from transformers import (BitsAndBytesConfig, WhisperForConditionalGeneration,
+                          WhisperProcessor)
+import torch
+
+# keep the regime the adapter was selected under: 4-bit NF4 base, NOT merged, bf16 autocast
+bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
+                         bnb_4bit_use_double_quant=True,
+                         bnb_4bit_compute_dtype=torch.bfloat16)
+base = WhisperForConditionalGeneration.from_pretrained(
+    "openai/whisper-large-v3", quantization_config=bnb, dtype=torch.bfloat16)
+model = PeftModel.from_pretrained(base, "yehoshua01/waxal-qlora-largev3-lin").eval()
+proc = WhisperProcessor.from_pretrained("openai/whisper-large-v3")
+```
+
+## The rest of the system
+
+| | |
+|---|---|
+| code, method, verification | [yehoshua0/waxal-asr-phase2](https://github.com/yehoshua0/waxal-asr-phase2) |
+| cached decodes and chain inputs | [`yehoshua01/waxal-phase2-chain-inputs`](https://huggingface.co/datasets/yehoshua01/waxal-phase2-chain-inputs) |
+| all checkpoints | [`yehoshua01` on the Hub](https://huggingface.co/yehoshua01) |
+
+**Sibling checkpoints** (primaries, voters and ablations of the same system): [`waxal-mms-1b-lin-pl2-spk`](https://huggingface.co/yehoshua01/waxal-mms-1b-lin-pl2-spk) · [`waxal-sunbird51-sna-pl2-spk`](https://huggingface.co/yehoshua01/waxal-sunbird51-sna-pl2-spk) · [`waxal-whisper-turbo-lin-r1`](https://huggingface.co/yehoshua01/waxal-whisper-turbo-lin-r1) · [`waxal-whisper-turbo-lin-r2`](https://huggingface.co/yehoshua01/waxal-whisper-turbo-lin-r2) · [`waxal-omni-ctc1b-lin`](https://huggingface.co/yehoshua01/waxal-omni-ctc1b-lin) · [`waxal-omni-ctc1b-sna`](https://huggingface.co/yehoshua01/waxal-omni-ctc1b-sna) · [`waxal-sunbird51-lin-ft-r2`](https://huggingface.co/yehoshua01/waxal-sunbird51-lin-ft-r2) · [`waxal-sunbird51-lin-ft-light`](https://huggingface.co/yehoshua01/waxal-sunbird51-lin-ft-light) · [`waxal-mms-1b-lin-full`](https://huggingface.co/yehoshua01/waxal-mms-1b-lin-full) · [`waxal-mms-1b-lin-fullmeta`](https://huggingface.co/yehoshua01/waxal-mms-1b-lin-fullmeta) · [`waxal-ssa-hubert-lin`](https://huggingface.co/yehoshua01/waxal-ssa-hubert-lin)
+
+
+## Licence and intended use
+
+`apache-2.0`. Training data is [`google/WaxalNLP`](https://huggingface.co/datasets/google/WaxalNLP)
+(CC-BY-SA-4.0, share-alike), so derivatives carry that too.
+
+**These weights are not a general-purpose ASR model.** Pseudo-labels were computed on the phase-2
+test audio (transductive self-training, permitted for phase-2 training by the host), so the
+checkpoint is partly adapted to that specific set.