Commit History
Today
-
docs: publish the verified evaluation stats
Luis J CamargoReplaces the "planned" placeholder with the results of the independent verification run against the Hub copy at 8085758ff8 (17 checks passed, 4 warnings, 2 skipped), generated from the report tables. - Global performance: top-1, top-3, macro-F1, macro recall, family and superlanguage accuracy with 95 % intervals (Wilson / bootstrap), next to the validation figures. Reproduced values differ from metrics.json by <= 0.002. - Per family: variants, test clips, top-1 and family accuracy. - Best and worst performing variants: highest recall with >= 20 test clips, lowest recall with >= 5, each with its interval and main confusion; recall bands, medians and the 0.174 Spearman correlation between training size and recall. - Confusions: 64.1 % of errors stay inside the correct family, top 10 pairs flagged same-family or not. - Confidence and clip length: selective accuracy (0.906 at full coverage up to 0.974 at 81.5 % coverage) and accuracy by duration. - Optimism diagnostics: ch…ec90e8f -
docs: tighten the header and the Tachiwin closing paragraph
Luis J Camargo- Icon moved above everything (it is the first element of the card). - Subtitle no longer states a class count: it read as "128 variants exist in Mexico". - Intro paragraph rewritten in a single homogeneous technical register: what the task is, that the label space is mostly ISO 639-3 with a few catalogue-specific additions, that it does not transcribe, that the ECAPA-TDNN targets an in-device distributable model, and that it ships with its configuration and per-language statistics. Dropped the "small on purpose" framing and the repeated class count as a selling point. - Closing Tachiwin paragraph: general mission (AI technologies for the 68 indigenous languages and their 364 variants) instead of the identify/translate/converse roadmap, and the funding sentence removed. - Removed the acknowledgement line; the obituary is kept. - Nothing else touched: Usage, Model details, Stats, Training, Limitations and Citation are unchanged.bcef18d -
docs: tighten the model card to an archetypal README
Luis J Camargo- Structure: title, one-line subtitle, icon, a single intro paragraph (what the model is, with only brief references to the project's aims), then Usage, Model details and architecture, Stats, Training, Limitations, Citation and the Tachiwin paragraph at the end. - Removed the sections that were editorial additions rather than source material: "Why this model exists", "What it does" (its table is gone; the aims are prose and the specs stay in the spec table), "Applications", "Responsible use", "How we work", the planned-stats block and the two appendices listing all 128 classes. - Card went from 566 to 293 lines (128 of which are the frontmatter language list); the Spanish version from 402 to 143. - No technical detail dropped: base model, dataset revision, temperature, windowing, every validation/test metric, per-language pointers, the training configuration and all limitations are still there. - Icon is now referenced relatively (`icon.png`) instead of an absolute …b6b40a8 -
docs: rewrite the model card, add Spanish version, rename to Tachiwin-Audio-ID
Luis J Camargo- README.md rewritten around the project's motives, aims and ethics (Tachiwin is indigenous, Mexican, community-led, non-profit and open source), while keeping every technical detail of the previous card and adding: task/IO spec, architecture and calibration table, repository file map, training data and procedure, full validation + test metrics, top confusions, scope of use, responsible-use guidance, citation and per-variant appendix (all 128 classes with INALI name, English name, superlanguage and family). - Frontmatter: pretty_name, all 128 ISO 639-3 codes in `language`, extra tags, and a commented model-index scaffold for the Hub eval widget. - README.es.md: full Spanish version, linked both ways, house voice. - Repository id moved to tachiwin/Tachiwin-Audio-ID in both cards, in the inference.py usage docstring and in train_config.json FINAL_REPO. The -ckpt repository is untouched. - icon.png added for branding consistency with the demo Space.1edf2f7
Yesterday
-
data: sync language names with the corrected INALI catalogue
Luis J CamargoThe `language` field in labels.json and lang_stats.json now uses the catalogue's INALI name, with spanish_name as fallback (which is what the app already displays for the Spanish UI). - 119 of 128 names updated; 9 already matched. - Mostly the Title Case normalization, plus a few real corrections (e.g. chq "Chinanteco de la Sierra de Juárez" -> "Chinanteco de Quiotepec", stp "Tepehuano del Sureste" -> "Tepehuano del Sur Bajo", pua/tsz "Purepecha" -> "Purépecha", zar "de Rincón" -> "del Rincón"). - nah is absent from the catalogue and keeps its existing name. - Nothing else touched: class order, superlanguage, family and all statistics are unchanged.8085758
2 days ago
-
final export (step 15210, best step 12168)
Verified signature46b9620 -
initial commit
Verified signature66402c0
