Datasets:
README: rewrite intro to frame as RULER-inspired, explain near-duplicate task properly
Browse files
README.md
CHANGED
|
@@ -109,16 +109,30 @@ configs:
|
|
| 109 |
|
| 110 |
# Genome-NIAH
|
| 111 |
|
| 112 |
-
A
|
| 113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
|
| 115 |
## Tasks
|
| 116 |
|
| 117 |
| Task | What it tests |
|
| 118 |
|---|---|
|
| 119 |
-
| `niah` | Plain NIAH: insert (KEY, VALUE) at a
|
| 120 |
-
| `niah_neardup_d4` | NIAH + 8 distractor (KEY', VALUE') pairs where KEY' differs from KEY by 4 bp (83 % key identity). |
|
| 121 |
-
| `niah_neardup_d2` | Same but Δ = 2 bp (92 % key identity). |
|
| 122 |
| `niah_neardup_d1` | Same but Δ = 1 bp (96 % key identity) — the discrimination test. |
|
| 123 |
|
| 124 |
## Context lengths
|
|
|
|
| 109 |
|
| 110 |
# Genome-NIAH
|
| 111 |
|
| 112 |
+
A long-context retrieval benchmark for genomic language models, **inspired by
|
| 113 |
+
NIAH and RULER**.
|
| 114 |
+
|
| 115 |
+
The plain `niah` task is a direct DNA analogue of RULER's `niah_single`:
|
| 116 |
+
insert a random (KEY, VALUE) pair into a long real-genome haystack, append
|
| 117 |
+
KEY at the end, and ask the model to generate VALUE. This gives a clean
|
| 118 |
+
cliff diagnostic — the context length at which a model stops retrieving.
|
| 119 |
+
|
| 120 |
+
To stress-test models that already pass plain NIAH, we developed a harder
|
| 121 |
+
**near-duplicate** variant. Inspired by RULER's `niah_multikey` (random
|
| 122 |
+
distractor pairs), we replace random distractors with **8 (KEY', VALUE')
|
| 123 |
+
pairs whose KEY' differs from the target KEY by exactly Δ base pairs**
|
| 124 |
+
(Hamming distance Δ = 1, 2, or 4). The model must distinguish the target
|
| 125 |
+
key from 8 close-but-not-identical decoys — a fine-grained pattern-matching
|
| 126 |
+
test that has biological relevance (DNA contains many similar-but-distinct
|
| 127 |
+
motifs) and provides resolution beyond what plain NIAH can detect.
|
| 128 |
|
| 129 |
## Tasks
|
| 130 |
|
| 131 |
| Task | What it tests |
|
| 132 |
|---|---|
|
| 133 |
+
| `niah` | Plain NIAH: insert (KEY, VALUE) at a depth, query KEY → expect VALUE. The cliff diagnostic. |
|
| 134 |
+
| `niah_neardup_d4` | NIAH + 8 distractor (KEY', VALUE') pairs where KEY' differs from KEY by 4 bp (83 % key identity). Easiest harder tier. |
|
| 135 |
+
| `niah_neardup_d2` | Same but Δ = 2 bp (92 % key identity). Cleanest separator between extension methods. |
|
| 136 |
| `niah_neardup_d1` | Same but Δ = 1 bp (96 % key identity) — the discrimination test. |
|
| 137 |
|
| 138 |
## Context lengths
|