Add dataset card

README.md

@@ -1,51 +1,18 @@
 ---
 license: cc-by-4.0
 task_categories:
-- question-answering
+  - question-answering
 language:
-- en
+  - en
 tags:
-- benchmark
-- evaluation
-- letter-counting
-- tokenization
-- reasoning
+  - benchmark
+  - evaluation
+  - letter-counting
+  - tokenization
+  - reasoning
 size_categories:
-- n<1K
+  - n<1K
 pretty_name: StrawberryBench
-dataset_info:
-  features:
-  - name: id
-    dtype: string
-  - name: word
-    dtype: string
-  - name: letter
-    dtype: string
-  - name: question
-    dtype: string
-  - name: answer
-    dtype: int64
-  - name: difficulty
-    dtype: string
-  - name: word_length
-    dtype: int64
-  - name: zero_count
-    dtype: bool
-  - name: template_idx
-    dtype: int64
-  - name: language
-    dtype: string
-  splits:
-  - name: test
-    num_bytes: 130887
-    num_examples: 847
-  download_size: 29895
-  dataset_size: 130887
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
 ---
 
 # StrawberryBench Dataset
@@ -89,32 +56,36 @@ English (en)
 | `id` | string | Unique example identifier (`sb_XXXXX`) |
 | `word` | string | The word or phrase to count in |
 | `letter` | string | The single letter to count (lowercase) |
-| `question` | string | Natural language question (one of 4 templates) |
+| `question` | string | Natural language question |
 | `answer` | int | Correct count (ground truth) |
-| `difficulty` | string | `easy` / `medium` / `hard` / `sentence` |
+| `difficulty` | string | `easy` / `medium` / `hard` / `sentence` / `paragraph` / `names` / `foreign` |
 | `word_length` | int | Number of non-space characters in `word` |
 | `zero_count` | bool | Whether `answer == 0` (letter absent) |
-| `template_idx` | int | Which question template was used (0–3) |
+| `template_idx` | int | Which question template was used (0 or 1) |
+| `language` | string | For `foreign` difficulty, the language (optional) |
 
 ### Data Splits
 
 | Split | Examples |
 |-------|----------|
 | train | — |
-| test | 560 |
+| test | 847 |
 
 The dataset is released as a single test split only; it is designed for zero-shot / few-shot evaluation, not fine-tuning.
 
 ### Difficulty Tiers
 
-| Tier | Word Length | # Examples |
+| Tier | Description / Word Length | # Examples |
 |------|-------------|------------|
-| easy | 3–6 chars | 185 |
-| medium | 7–12 chars | 204 |
-| hard | 13+ chars | 132 |
-| sentence | phrase | 39 |
+| easy | 3–6 chars | 183 |
+| medium | 7–12 chars | 228 |
+| hard | 13+ chars | 146 |
+| sentence | short multi-word phrases | 46 |
+| paragraph | longer text passages (100+ chars) | 55 |
+| names | common first names (repeated letters) | 112 |
+| foreign | non-English words (German, etc.) | 77 |
 
-Approximately 25% of examples have `zero_count = true` (the letter does not appear in the word).
+Approximately 11% of examples have `zero_count = true` (the letter does not appear in the word).
 
 ## Dataset Creation
 
@@ -124,14 +95,15 @@ All words and phrases are common English vocabulary, proper nouns, and well-know
 
 ### Question Templates
 
-Four templates are used per question (randomly assigned):
+Two main templates are used per example (randomly assigned index 0 or 1):
 
-1. `"How many times does the letter '{letter}' appear in the word '{word}'?"`
-2. `"Count the occurrences of the letter '{letter}' in '{word}'."`
-3. `"In the word '{word}', how many '{letter}'s are there?"`
-4. `"What is the count of the letter '{letter}' in '{word}'?"`
-
-Sentence-level questions use analogous templates with "phrase" instead of "word".
+- **Standard (easy, medium, hard):**
+  1. `"How many times does the letter '{letter}' appear in the word '{word}'?"`
+  2. `"In the word '{word}', how many '{letter}'s are there?"`
+- **Sentences:** Uses "phrase" instead of "word".
+- **Paragraphs:** Uses "following text: '{word}'".
+- **Names:** `"How many {letter}'s are in the name '{word}'?"`
+- **Foreign:** `"How many times does the letter '{letter}' appear in the {language} word '{word}'?"`
 
 ### Curation Rationale
 
@@ -151,17 +123,6 @@ correct = sum(1 for ex in ds["test"] if model_predict(ex["question"]) == ex["ans
 accuracy = correct / len(ds["test"])
 ```
 
-## Citation
-
-```bibtex
-@article{strawberrybench2026,
-  title={StrawberryBench: Can Language Models Count Letters?},
-  author={},
-  journal={arXiv preprint arXiv:XXXX.XXXXX},
-  year={2026}
-}
-```
-
 ## License
 
 [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)