Update README.md

README.md

@@ -16,23 +16,17 @@ pipeline_tag: text2text-generation
 
 # byt5-mitra-bo-tagger
 
-A byte-level seq2seq tagger for **Classical Tibetan in Wylie transliteration**, aimed at learners of
-Tibetan in Buddhist studies who read Tibetan translations against their Sanskrit originals, and at
-dictionary building. Given a block of raw Wylie it returns, in one terse string:
+A byte-level seq2seq tagger for Classical Tibetan in Wylie transliteration. Given a block of raw Wylie it returns, in one terse string:
 
 1. **sentence segmentation** (one sentence per line),
-2. **dictionary-lexeme word segmentation** (units a learner looks up in Hopkins / Rangjung Yeshe / Negi:
-   `sems_can`, `byang_chub_sems_dpa'`, `tshe_dang_ldan_pa` stay whole; case particles, clause particles,
-   clitics and attached affixes are separate),
+2. **dictionary-lexeme word segmentation** 
 3. a **coarse part of speech** for every word (noun / verb / verbal noun / adjective / particle …), and
 4. **Sanskrit-unit brackets**: runs of Tibetan words that render one Sanskrit word or compound
    (`[chos/N kyi/C rgyal_po/N]` = dharmarāja).
 
 It is a fine-tune of [buddhist-nlp/byt5-mitra-bo](https://huggingface.co/buddhist-nlp/byt5-mitra-bo)
 (ByT5 continued-pretrained on Tibetan Wylie) on ~5k sentences and short paragraphs from the
-Sanskrit–Tibetan parallel corpus of the [Dharmamitra](https://dharmamitra.org) project. The training
-annotations were produced by Gemini from the Tibetan, the Sanskrit original with a morphological
-analysis, and automatic Sanskrit–Tibetan word alignments, then validated to reproduce the input exactly.
+Sanskrit–Tibetan parallel corpus of the [Dharmamitra](https://dharmamitra.org) project. 
 
 ## Output scheme
 
@@ -50,9 +44,6 @@ POS letters: **N** noun · **P** proper noun · **V** verb (finite / predicate s
 (demonstrative, plural, indefinite) · **X** negation · **C** case particle · **K** clause particle
 (after verbs) · **L** clitic / final particle · **I** interjection · **p** punctuation.
 
-The surface text is recoverable (`_` → space, drop `+`, `|` → `/`), so outputs can be validated
-against the input; ~10% of outputs fail that check (a dropped or altered syllable) and should be
-rejected or retried.
 
 ```
 input : de nas tshe dang ldan pa kun dga' bos bcom ldan 'das la 'di skad ces gsol to/ /
@@ -84,15 +75,13 @@ Inputs: Wylie (EWTS) as found in ACIP / DharmaNexus e-texts, up to ~700 characte
 
 ## Evaluation
 
-Held-out 100 items (reference = the Gemini teacher annotation):
+Held-out 100 items:
 
 | model | valid output | word boundary F1 | POS acc | verb/noun acc | sentence F1 | Sanskrit-unit F1 |
 |---|---|---|---|---|---|---|
 | byt5-mitra-bo-tagger (this model, 10 epochs) | .90 | .959 | .914 | .914 | .914 | .663 |
 | Qwen3.5-9B mitra base, same data | 1.00 | .962 | .942 | .909 | .891 | .687 |
 
-Two runs of the teacher itself agree at word F1 .974 / POS .947, so the tagger is close to the
-teacher's own consistency on segmentation.
 
 ## Limitations