Skip to content

Commit 5fdcd44

Browse files
committed
Add TEX-T
1 parent df64ecf commit 5fdcd44

File tree

5 files changed

+19
-17
lines changed

5 files changed

+19
-17
lines changed

docs/approaches.csv

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -13,4 +13,5 @@ AutoCodeRover,https://autocoderover.dev,autocoderover,https://github.com/logic-s
1313
LIBRO,https://arxiv.org/abs/2209.11515,kaist,https://github.com/logic-star-ai/swt-bench?tab=readme-ov-file#evaluation-results
1414
Otter++,https://arxiv.org/abs/2502.05368v1,ibm,https://files.sri.inf.ethz.ch/swt-bench/otter/
1515
Otter,https://arxiv.org/abs/2502.05368v1,ibm,https://files.sri.inf.ethz.ch/swt-bench/otter/
16-
LogicStar AI,https://logicstar.ai/,logicstar,https://github.com/logic-star-ai/swt-bench?tab=readme-ov-file#evaluation-results
16+
LogicStar AI,https://logicstar.ai/,logicstar,https://github.com/logic-star-ai/swt-bench?tab=readme-ov-file#evaluation-results
17+
TEX-T,https://akashgokul.github.io/TEX/,salesforce,https://github.com/logic-star-ai/swt-bench?tab=readme-ov-file#evaluation-results

docs/orgs.csv

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9,4 +9,5 @@ swe-agent,SWE-agent,https://swe-agent.com/,./static/images/logos/swe-agent.svg
99
aider,Aider,https://aider.chat/,./static/images/logos/aider.png
1010
autocoderover,AutoCodeRover,https://autocoderover.net,./static/images/logos/autocoderover.svg
1111
kaist,KAIST,https://github.com/coinse/libro,./static/images/logos/KAIST.svg
12-
openai,OpenAI,https://openai.com,./static/images/logos/openai.svg
12+
openai,OpenAI,https://openai.com,./static/images/logos/openai.svg
13+
salesforce,Salesforce,https://www.salesforce.com/,./static/images/logos/Salesforce.com_logo.png

docs/runs.csv

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -27,4 +27,5 @@ verified,,Otter,GPT-4o,31.6,37.6,2025-03-10,unittest
2727
verified,,OpenHands,Cl. Sonnet 3.5,27.7,52.9,2025-02-28,unittest
2828
verified,,LIBRO,GPT-4o,17.8,38.0,2025-02-28,unittest
2929
verified,,Zero-Shot Plus,GPT-4o + BM25,14.3,34.0,2025-02-28,unittest
30-
verified,new,LogicStar AI,L*Agent v1, 84.0, 67.7,2025-09-13,unittest
30+
verified,new,LogicStar AI,L*Agent v1, 84.0, 67.7,2025-09-13,unittest
31+
verified,new,TEX-T,Claude 4 Sonnet,87.0,69.8,2025-12-17,reproduction
42.2 KB
Loading

docs/static/js/swtbench.js

Lines changed: 13 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -132,21 +132,20 @@ document.addEventListener("DOMContentLoaded", () => {
132132
document.addEventListener('DOMContentLoaded', function () {
133133
const reproductionCheckbox = document.getElementById('reproduction-mode-checkbox');
134134
const nonReproductionCheckbox = document.getElementById('unittest-mode-checkbox');
135-
const rows = document.querySelectorAll('#leaderboard-table tbody tr');
135+
const tableSelectors = [
136+
'#leaderboard-table tbody tr',
137+
'#verified-leaderboard-table tbody tr',
138+
];
136139

137140
function updateRowVisibility() {
138-
rows.forEach(row => {
139-
const mode = row.getAttribute('data-mode');
140-
if (
141-
(mode === 'reproduction' && !reproductionCheckbox.checked) ||
142-
(mode !== 'reproduction' && !nonReproductionCheckbox.checked)
143-
) {
144-
// Hide the row
145-
row.style.display = 'none';
146-
} else {
147-
// Show the row
148-
row.style.display = '';
149-
}
141+
tableSelectors.forEach(selector => {
142+
document.querySelectorAll(selector).forEach(row => {
143+
const mode = row.getAttribute('data-mode');
144+
const shouldShow =
145+
(mode === 'reproduction' && reproductionCheckbox.checked) ||
146+
(mode !== 'reproduction' && nonReproductionCheckbox.checked);
147+
row.style.display = shouldShow ? '' : 'none';
148+
});
150149
});
151150
}
152151

@@ -217,4 +216,4 @@ document.addEventListener('DOMContentLoaded', function () {
217216
// update upon loading the page
218217
updateTable(document.getElementById('leaderboard-table'));
219218
updateTable(document.getElementById('verified-leaderboard-table'));
220-
});
219+
});

0 commit comments

Comments
 (0)