olaverse
/

diacnet-1.1

@@ -12,7 +12,7 @@ language:
 - fr
 - it
 library_name: transformers
-pipeline_tag: text2text-generation
 base_model: google/byt5-base
 tags:
 - diacritics
@@ -26,6 +26,13 @@ tags:
 datasets:
 - olaverse/diacnet-1.1-corpus
 - olaverse/diacbench
 metrics:
 - cer
 - wer
@@ -62,7 +69,7 @@ print(gen("<pol> Lodz jest piekna"))            # Łódź jest piękna
 Evaluated on [diacbench](https://huggingface.co/datasets/olaverse/diacbench) — 1,000 held-out sentences per language, sourced independently of the training corpus. Metrics: **DER** (diacritic error rate, errors restricted to diacritic-eligible characters), **WER**, **CER**, exact sentence match. All figures below use n=300 per language with greedy decoding.
-![DER by language](./diacbench_der.png)
 | lang | DER | WER | CER | exact | DER with `<auto>` |
 |---|---:|---:|---:|---:|---:|
@@ -94,7 +101,7 @@ Low exact-match on Vietnamese and Yorùbá reflects how many marks those languag
 | hau | **0.0383** | 0.0593 | 8.58% | **2.53%** | mixed |
 | yor | **0.1554** | 0.2006 | **27.69%** | 42.18% | 1.0 better |
-![WER by language](./diacbench_wer.png)
 Note that DER and WER disagree on ibo/hau/fra: v1.1 makes *fewer wrong words* but a higher share of diacritic-level errors within them. For Hausa the difference is stark — v1.0 got some detail wrong in almost every sentence (0% exact match), while v1.1 is perfect on 60% of sentences and worse on the rest.
@@ -134,7 +141,7 @@ Scored strictly, **diacnet-1.1 beats Claude on 8 of 10 languages**. With fallbac
 | Claude 4.5 | 74.3% | 79.3% | 65.7% | 97.7% | 96.7% | 98.0% | 97.7% | 98.3% | 88.3% | 100% |
 | GPT-4o-mini | 68.0% | 57.3% | 49.3% | 93.7% | 96.0% | 94.7% | 97.7% | 97.3% | 96.7% | 99.7% |
-![Output compliance](./diacbench_compliance.png)
 **WER**

 - fr
 - it
 library_name: transformers
+pipeline_tag: text-generation
 base_model: google/byt5-base
 tags:
 - diacritics
 datasets:
 - olaverse/diacnet-1.1-corpus
 - olaverse/diacbench
+widget:
+- text: "<yor> se eranko naa si gbo o?"
+  example_title: "Yorùbá"
+- text: "<vie> Toi khong biet tieng Viet"
+  example_title: "Vietnamese"
+- text: "<auto> El nino esta en la casa"
+  example_title: "auto-detect"
 metrics:
 - cer
 - wer
 Evaluated on [diacbench](https://huggingface.co/datasets/olaverse/diacbench) — 1,000 held-out sentences per language, sourced independently of the training corpus. Metrics: **DER** (diacritic error rate, errors restricted to diacritic-eligible characters), **WER**, **CER**, exact sentence match. All figures below use n=300 per language with greedy decoding.
+![DER by language](https://cdn-uploads.huggingface.co/production/uploads/69949cbacd82af728f850c12/zVLYyvaWnrHx78T5k_tfA.png)
 | lang | DER | WER | CER | exact | DER with `<auto>` |
 |---|---:|---:|---:|---:|---:|
 | hau | **0.0383** | 0.0593 | 8.58% | **2.53%** | mixed |
 | yor | **0.1554** | 0.2006 | **27.69%** | 42.18% | 1.0 better |
+![WER by language](https://cdn-uploads.huggingface.co/production/uploads/69949cbacd82af728f850c12/zP3g59zYF0uGx6Hm-L-u_.png)
 Note that DER and WER disagree on ibo/hau/fra: v1.1 makes *fewer wrong words* but a higher share of diacritic-level errors within them. For Hausa the difference is stark — v1.0 got some detail wrong in almost every sentence (0% exact match), while v1.1 is perfect on 60% of sentences and worse on the rest.
 | Claude 4.5 | 74.3% | 79.3% | 65.7% | 97.7% | 96.7% | 98.0% | 97.7% | 98.3% | 88.3% | 100% |
 | GPT-4o-mini | 68.0% | 57.3% | 49.3% | 93.7% | 96.0% | 94.7% | 97.7% | 97.3% | 96.7% | 99.7% |
+![Output compliance](https://cdn-uploads.huggingface.co/production/uploads/69949cbacd82af728f850c12/eQcOc6vpurneX5j8VOydW.png)
 **WER**