Public text-only card: DOSE v1.0, 2466 rows, no audio

README.md

@@ -1,73 +1,42 @@
 ---
 license: cc-by-4.0
 task_categories:
-- text-to-speech
-- audio-classification
+  - text-to-speech
 language:
-- en
+  - en
 tags:
-- text-to-speech
-- tts
-- pronunciation
-- healthcare
-- pharma
-- drug-names
-- benchmark
-pretty_name: DOSE Benchmark (Drug-name Oral Synthesis Evaluation)
+  - text-to-speech
+  - tts
+  - pronunciation
+  - healthcare
+  - pharma
+  - drug-names
+  - benchmark
+pretty_name: DOSE v1.0 (Drug-name Oral Synthesis Evaluation)
 size_categories:
-- 1K<n<10K
+  - 1K<n<10K
 configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: drug
-    dtype: string
-  - name: name_type
-    dtype:
-      class_label:
-        names:
-          '0': brand
-          '1': generic
-  - name: sentence
-    dtype: string
-  - name: ground_truth_ipa
-    dtype: string
-  - name: system_key
-    dtype: string
-  - name: system_name
-    dtype: string
-  - name: heard_ipa
-    dtype: string
-  - name: score
-    dtype: int32
-  - name: passed
-    dtype: bool
-  - name: has_audio
-    dtype: bool
-  splits:
-  - name: train
-    num_bytes: 598164
-    num_examples: 2466
-  download_size: 54195
-  dataset_size: 598164
+  - config_name: default
+    data_files:
+      - split: train
+        path: data/train-*.parquet
 ---
 
-# DOSE v1: Drug-name Oral Synthesis Evaluation
+# DOSE v1.0: Drug-name Oral Synthesis Evaluation
 
 **Can your voice model say retatrutide?**
 
 DOSE is a benchmark measuring how accurately text-to-speech (TTS) systems
-pronounce pharmaceutical drug names. It covers both well-established brand
-names and newly-approved names (many with no training-data exposure), as
-well as generic/INN names, each tested inside a real clinical sentence
-rather than spoken in isolation.
+pronounce pharmaceutical drug names. It covers established brands,
+newly-approved names, and generic/INN names, each tested inside a real
+clinical sentence rather than spoken in isolation.
 
-Leaderboard and five public audio samples: https://www.synthiolabs.com/synthio-dose-benchmark
+**Version: DOSE v1.0.** Live leaderboard and five public audio samples:
+https://www.synthiolabs.com/synthio-dose-benchmark
 
-**Version: DOSE v1.0.** The full dataset is gated on Hugging Face; request access to download it.
+This public dataset is **text-only** (sentences, reference pronunciations,
+per-system scores). The evaluation audio is not released here. Request the
+full audio set at [rohit@synthiolabs.com](mailto:rohit@synthiolabs.com?subject=DOSE%20full%20audio%20request).
 
 ## Why this matters
 
@@ -87,10 +56,9 @@ ds = load_dataset("SynthioLabs/dose-benchmark", split="train")
 
 row = ds[0]
 print(row["drug"], row["system_name"], row["score"], row["passed"])
-row["audio"]  # decoded automatically -> {"array": np.ndarray, "sampling_rate": int}
 ```
 
-Filter to just one system's results, or just failures:
+Filter to one system, failures, or generic names:
 
 ```python
 synthio_only = ds.filter(lambda r: r["system_key"] == "synthio_rxpronounce")
@@ -98,7 +66,7 @@ failures = ds.filter(lambda r: not r["passed"])
 generic_names = ds.filter(lambda r: r["name_type"] == "generic")
 ```
 
-Recompute the leaderboard yourself:
+Recompute the leaderboard:
 
 ```python
 import pandas as pd
@@ -118,26 +86,25 @@ print(leaderboard)
 
 - **274 drug names**: 128 traditional (established) + 146 new (recently
   approved) names.
-- **Brand and generic/INN names** (`name_type` column): invented brand
-  names are designed to be pronounceable; generic names (e.g.
-  "acoramidis") are built from technical drug-naming stems and are not —
-  every system's accuracy drops on this cohort.
-- **9 TTS systems** evaluated on each name, each using its current
-  flagship model: ElevenLabs, OpenAI, Google (Gemini), Deepgram,
-  Microsoft Azure, Cartesia (two model versions), xAI, and Synthio
-  RxPronounce.
-- **2,461 scored utterances** total (one row per drug x system, minus a
-  handful of deterministic provider-side generation failures).
+- **Brand and generic/INN names** (`name_type` column).
+- **9 TTS systems** evaluated on each name.
+- **2,466 scored utterances** (274 names × 9 systems). Five of those are
+  provider-side generation failures recorded as score 0 with no audio.
 - Every system reads the **exact same clinical sentence** containing the
   drug name, with no special pronunciation hints given to any system.
 
-## Reference pronunciations
+## How DOSE is scored
 
-Each drug’s reference comes from public pharmaceutical sources, including labels, manufacturer materials, and established medical references. It is then checked against spoken clinical usage so the target reflects the accepted pronunciation.
+**Reference pronunciations.** Each drug’s reference comes from public
+pharmaceutical sources, including labels, manufacturer materials, and
+established medical references. It is then checked against spoken clinical
+usage so the target reflects the accepted pronunciation — not a model’s
+guess.
 
-## Scoring
-
-Each system’s recording is scored 0–5 against the same reference by automated audio comparison. Several independent judgments are combined into one result; **4 or higher counts as a pass** (`passed` column). **Human reviewers are not in the scoring loop.**
+**Judge.** Each system’s recording is scored 0–5 against that same
+reference by automated audio comparison. Several independent judgments are
+combined into one result; **4 or higher counts as a pass**. Human
+reviewers are not in the scoring loop.
 
 ## Columns
 
@@ -145,14 +112,14 @@ Each system’s recording is scored 0–5 against the same reference by automate
 |---|---|---|
 | `drug` | string | Drug name (brand or generic/INN) |
 | `name_type` | class label | `brand` or `generic` |
-| `sentence` | string | The clinical sentence read aloud, containing the drug name |
+| `sentence` | string | The clinical sentence read aloud |
 | `ground_truth_ipa` | string | Correct pronunciation (informal phonetic spelling) |
 | `system_key` | string | Machine-readable system identifier |
 | `system_name` | string | Human-readable TTS system name |
-| `heard_ipa` | string | Phonetic transcription of what the system actually said |
-| `score` | int (0-5) | Judged pronunciation accuracy |
+| `heard_ipa` | string | Phonetic transcription of what the system said |
+| `score` | int (0–5) | Judged pronunciation accuracy |
 | `passed` | bool | `score >= 4` |
-| `audio` | audio | The generated utterance for this (drug, system) pair |
+| `has_audio` | bool | False for the five provider-side generation failures |
 
 ## Leaderboard (overall pass rate, score >= 4)
 
@@ -170,13 +137,14 @@ Each system’s recording is scored 0–5 against the same reference by automate
 
 ## Submit your model
 
-Submit your model for evaluation at [rohit@synthiolabs.com](mailto:rohit@synthiolabs.com?subject=DOSE%20model%20submission).
+Email us at [rohit@synthiolabs.com](mailto:rohit@synthiolabs.com?subject=DOSE%20model%20submission)
+to submit a model for evaluation, or to request the full evaluation audio.
 
 ## Citation
 
 ```bibtex
 @misc{synthio2026dosev1,
-  title   = {DOSE v1: Drug-name Oral Synthesis Evaluation},
+  title   = {DOSE v1.0: Drug-name Oral Synthesis Evaluation},
   author  = {Pari, Rohit Gangadhar and {Synthio Labs}},
   year    = {2026},
   url     = {https://www.synthiolabs.com/synthio-dose-benchmark}