nkundiushuti commited on
Commit
c3d05c2
·
verified ·
1 Parent(s): 64c2b71

Add BEANS-Next dataset card (tier/subset layout)

Browse files
Files changed (1) hide show
  1. README.md +63 -0
README.md ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ configs: []
4
+ ---
5
+
6
+ # beans-next
7
+
8
+ **beans-next** is the Hugging Face dataset for the **BEANS-Next** Parquet bundle: metadata shards that mirror the public JSONL splits. Audio is **not** embedded—each row keeps GCS URIs so loaders fetch WAV bytes the same way as the `esp_data` path.
9
+
10
+ **Tiers 1-3** (directories ``tier_1``, ``tier_2``, ``tier_3``) hold only single-audio **Beans-Next** subsets. **Tier 4** (``tier_4_in_context``) holds **only** **Beans-Next** in-context (multi-audio) subsets; those subsets do not appear in tiers 1-3.
11
+
12
+ ## Repository layout
13
+
14
+ Each subset has one default shard (``test`` split only):
15
+
16
+ - **tier_1/`<subset>`/test.parquet** — tier 1, Beans-Next (single-audio)
17
+ - **tier_2/`<subset>`/test.parquet** — tier 2, Beans-Next (single-audio)
18
+ - **tier_3/`<subset>`/test.parquet** — tier 3, Beans-Next (single-audio)
19
+ - **tier_4_in_context/`<subset>`/test.parquet** — tier 4 only, Beans-Next (in-context)
20
+
21
+ ### Tier 1 subsets
22
+
23
+ - `crow-description`
24
+ - `f0-mean-heldout-taxa`
25
+ - `f0-mean-seen-taxa`
26
+ - `zebra-description`
27
+
28
+ ### Tier 2 subsets
29
+
30
+ - `alarm-call-presence`
31
+ - `amphibian-presence`
32
+ - `bird-presence`
33
+ - `call-type-fixed-vocab`
34
+ - `flight-call-presence`
35
+ - `mammal-presence`
36
+
37
+ ### Tier 3 subsets
38
+
39
+ - `insect-presence`
40
+
41
+ ### Tier 4 subsets (Beans-Next in-context)
42
+
43
+ - `crow-4way`
44
+ - `dcase-fewshot-detection-balanced`
45
+ - `giant-otter-4way`
46
+ - `gibbon-fewshot-detection-balanced`
47
+ - `unseen-species-4way`
48
+ - `zebra-4way`
49
+
50
+ ## Provenance
51
+
52
+ Consolidated with ``scripts/consolidate_beans_next_parquet.py`` from the public GCS JSONL paths encoded for **Beans-Next** in ``beans_pro.datasets.beans_next_parquet``.
53
+
54
+ ## Loading with ``datasets``
55
+
56
+ Each subset is a separate **configuration** (``config_name`` == subset id). The Hub README maps the Parquet shard to split **``test``** (not ``train``).
57
+
58
+ ```python
59
+ from datasets import load_dataset
60
+
61
+ ds = load_dataset("EarthSpeciesProject/beans-next", "crow-description", split="test")
62
+ print(ds[0])
63
+ ```