Spaces:
Running
Running
Commit ·
ab75229
1
Parent(s): 81071cf
Fix open model filter callback reliability (#38)
Browse files- Fix open model filter callback reliability (5c02fd06b29fccf7afb556e9f4eb5d647f6ea33e)
- leaderboard_transformer.py +103 -100
- ui_components.py +9 -1
- visualizations.py +11 -1
leaderboard_transformer.py
CHANGED
|
@@ -12,6 +12,38 @@ from constants import FONT_FAMILY, FONT_FAMILY_SHORT
|
|
| 12 |
|
| 13 |
logger = logging.getLogger(__name__)
|
| 14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
# Company logo mapping for graphs - maps model name patterns to company logo files
|
| 16 |
COMPANY_LOGO_MAP = {
|
| 17 |
"anthropic": {"path": "assets/logo-anthropic.svg", "name": "Anthropic"},
|
|
@@ -109,42 +141,34 @@ def get_openhands_logo_images():
|
|
| 109 |
images = []
|
| 110 |
|
| 111 |
# Light mode logo (visible in light mode, hidden in dark mode)
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
yanchor="bottom",
|
| 126 |
-
))
|
| 127 |
-
except Exception:
|
| 128 |
-
pass
|
| 129 |
|
| 130 |
# Dark mode logo (hidden in light mode, visible in dark mode)
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
yanchor="bottom",
|
| 145 |
-
))
|
| 146 |
-
except Exception:
|
| 147 |
-
pass
|
| 148 |
|
| 149 |
return images
|
| 150 |
|
|
@@ -511,54 +535,50 @@ def create_scatter_chart(
|
|
| 511 |
marker_info = get_marker_icon(model_name, openness, mark_by)
|
| 512 |
logo_path = marker_info['path']
|
| 513 |
|
| 514 |
-
|
| 515 |
-
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
|
| 519 |
-
|
| 520 |
-
|
| 521 |
-
|
| 522 |
-
|
| 523 |
-
|
| 524 |
-
|
| 525 |
-
|
| 526 |
-
|
| 527 |
-
|
| 528 |
-
|
| 529 |
-
|
| 530 |
-
|
| 531 |
-
|
| 532 |
-
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
|
| 538 |
-
|
| 539 |
-
|
| 540 |
-
|
| 541 |
-
|
| 542 |
-
|
| 543 |
-
|
| 544 |
-
|
| 545 |
-
|
| 546 |
-
|
| 547 |
-
|
| 548 |
-
|
| 549 |
-
|
| 550 |
-
|
| 551 |
-
|
| 552 |
-
|
| 553 |
-
|
| 554 |
-
|
| 555 |
-
|
| 556 |
-
|
| 557 |
-
|
| 558 |
-
layer="above"
|
| 559 |
-
))
|
| 560 |
-
except Exception:
|
| 561 |
-
pass
|
| 562 |
|
| 563 |
# Add labels for frontier points only
|
| 564 |
for row in frontier_rows:
|
|
@@ -1193,25 +1213,8 @@ def _plot_scatter_plotly(
|
|
| 1193 |
y_min = min_score - 5 if min_score > 5 else 0
|
| 1194 |
y_max = max_score + 5
|
| 1195 |
|
| 1196 |
-
# Cache base64-encoded logos across rows — every Claude model on the
|
| 1197 |
-
# Alternative Agents page points at the same assets/harness-claude-code.svg,
|
| 1198 |
-
# so decoding once per path is ~N× cheaper than once per point.
|
| 1199 |
-
_logo_cache: dict[str, str] = {}
|
| 1200 |
def _encode_logo(path: str) -> Optional[str]:
|
| 1201 |
-
|
| 1202 |
-
return _logo_cache[path]
|
| 1203 |
-
if not os.path.exists(path):
|
| 1204 |
-
return None
|
| 1205 |
-
try:
|
| 1206 |
-
with open(path, "rb") as f:
|
| 1207 |
-
encoded = base64.b64encode(f.read()).decode("utf-8")
|
| 1208 |
-
except Exception as e:
|
| 1209 |
-
logger.warning(f"Could not load logo {path}: {e}")
|
| 1210 |
-
return None
|
| 1211 |
-
mime = "svg+xml" if path.lower().endswith(".svg") else "png"
|
| 1212 |
-
uri = f"data:image/{mime};base64,{encoded}"
|
| 1213 |
-
_logo_cache[path] = uri
|
| 1214 |
-
return uri
|
| 1215 |
|
| 1216 |
# Composite markers: on the Alternative Agents page the dataframe carries
|
| 1217 |
# an "Agent" column (Claude Code / Codex / Gemini CLI / OpenHands Sub-agents),
|
|
|
|
| 12 |
|
| 13 |
logger = logging.getLogger(__name__)
|
| 14 |
|
| 15 |
+
_DATA_URI_CACHE: dict[str, str] = {}
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def get_asset_data_uri(path: str) -> Optional[str]:
|
| 19 |
+
"""Return a cached data URI for a local image asset."""
|
| 20 |
+
if path in _DATA_URI_CACHE:
|
| 21 |
+
return _DATA_URI_CACHE[path]
|
| 22 |
+
|
| 23 |
+
if not os.path.exists(path):
|
| 24 |
+
_DATA_URI_CACHE[path] = ""
|
| 25 |
+
return None
|
| 26 |
+
|
| 27 |
+
try:
|
| 28 |
+
with open(path, "rb") as f:
|
| 29 |
+
encoded = base64.b64encode(f.read()).decode("utf-8")
|
| 30 |
+
except Exception as e:
|
| 31 |
+
logger.warning(f"Could not load image asset {path}: {e}")
|
| 32 |
+
_DATA_URI_CACHE[path] = ""
|
| 33 |
+
return None
|
| 34 |
+
|
| 35 |
+
ext = os.path.splitext(path)[1].lower()
|
| 36 |
+
if ext == ".svg":
|
| 37 |
+
mime = "image/svg+xml"
|
| 38 |
+
elif ext == ".png":
|
| 39 |
+
mime = "image/png"
|
| 40 |
+
else:
|
| 41 |
+
mime = "application/octet-stream"
|
| 42 |
+
|
| 43 |
+
uri = f"data:{mime};base64,{encoded}"
|
| 44 |
+
_DATA_URI_CACHE[path] = uri
|
| 45 |
+
return uri
|
| 46 |
+
|
| 47 |
# Company logo mapping for graphs - maps model name patterns to company logo files
|
| 48 |
COMPANY_LOGO_MAP = {
|
| 49 |
"anthropic": {"path": "assets/logo-anthropic.svg", "name": "Anthropic"},
|
|
|
|
| 141 |
images = []
|
| 142 |
|
| 143 |
# Light mode logo (visible in light mode, hidden in dark mode)
|
| 144 |
+
light_logo_uri = get_asset_data_uri(OPENHANDS_LOGO_PATH_LIGHT)
|
| 145 |
+
if light_logo_uri:
|
| 146 |
+
images.append(dict(
|
| 147 |
+
source=light_logo_uri.replace("data:image/png;base64,", "data:image/png;openhands=lightlogo;base64,"),
|
| 148 |
+
xref="paper",
|
| 149 |
+
yref="paper",
|
| 150 |
+
x=0,
|
| 151 |
+
y=-0.15,
|
| 152 |
+
sizex=0.15,
|
| 153 |
+
sizey=0.15,
|
| 154 |
+
xanchor="left",
|
| 155 |
+
yanchor="bottom",
|
| 156 |
+
))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
|
| 158 |
# Dark mode logo (hidden in light mode, visible in dark mode)
|
| 159 |
+
dark_logo_uri = get_asset_data_uri(OPENHANDS_LOGO_PATH_DARK)
|
| 160 |
+
if dark_logo_uri:
|
| 161 |
+
images.append(dict(
|
| 162 |
+
source=dark_logo_uri.replace("data:image/png;base64,", "data:image/png;openhands=darklogo;base64,"),
|
| 163 |
+
xref="paper",
|
| 164 |
+
yref="paper",
|
| 165 |
+
x=0,
|
| 166 |
+
y=-0.15,
|
| 167 |
+
sizex=0.15,
|
| 168 |
+
sizey=0.15,
|
| 169 |
+
xanchor="left",
|
| 170 |
+
yanchor="bottom",
|
| 171 |
+
))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 172 |
|
| 173 |
return images
|
| 174 |
|
|
|
|
| 535 |
marker_info = get_marker_icon(model_name, openness, mark_by)
|
| 536 |
logo_path = marker_info['path']
|
| 537 |
|
| 538 |
+
logo_uri = get_asset_data_uri(logo_path)
|
| 539 |
+
if not logo_uri:
|
| 540 |
+
continue
|
| 541 |
+
|
| 542 |
+
if x_type == "date":
|
| 543 |
+
# For date axes, use data coordinates directly
|
| 544 |
+
layout_images.append(dict(
|
| 545 |
+
source=logo_uri,
|
| 546 |
+
xref="x",
|
| 547 |
+
yref="y",
|
| 548 |
+
x=x_val,
|
| 549 |
+
y=y_val,
|
| 550 |
+
sizex=15 * 24 * 60 * 60 * 1000, # ~15 days in milliseconds
|
| 551 |
+
sizey=3, # score units
|
| 552 |
+
xanchor="center",
|
| 553 |
+
yanchor="middle",
|
| 554 |
+
layer="above"
|
| 555 |
+
))
|
| 556 |
+
else:
|
| 557 |
+
# For log axes, use domain coordinates (0-1 range)
|
| 558 |
+
if x_type == "log" and x_val > 0:
|
| 559 |
+
log_x = np.log10(x_val)
|
| 560 |
+
domain_x = (log_x - x_range_log[0]) / (x_range_log[1] - x_range_log[0])
|
| 561 |
+
else:
|
| 562 |
+
domain_x = 0.5
|
| 563 |
+
|
| 564 |
+
domain_y = (y_val - y_range[0]) / (y_range[1] - y_range[0]) if (y_range[1] - y_range[0]) > 0 else 0.5
|
| 565 |
+
|
| 566 |
+
# Clamp to valid range
|
| 567 |
+
domain_x = max(0, min(1, domain_x))
|
| 568 |
+
domain_y = max(0, min(1, domain_y))
|
| 569 |
+
|
| 570 |
+
layout_images.append(dict(
|
| 571 |
+
source=logo_uri,
|
| 572 |
+
xref="x domain",
|
| 573 |
+
yref="y domain",
|
| 574 |
+
x=domain_x,
|
| 575 |
+
y=domain_y,
|
| 576 |
+
sizex=0.04,
|
| 577 |
+
sizey=0.06,
|
| 578 |
+
xanchor="center",
|
| 579 |
+
yanchor="middle",
|
| 580 |
+
layer="above"
|
| 581 |
+
))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 582 |
|
| 583 |
# Add labels for frontier points only
|
| 584 |
for row in frontier_rows:
|
|
|
|
| 1213 |
y_min = min_score - 5 if min_score > 5 else 0
|
| 1214 |
y_max = max_score + 5
|
| 1215 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1216 |
def _encode_logo(path: str) -> Optional[str]:
|
| 1217 |
+
return get_asset_data_uri(path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1218 |
|
| 1219 |
# Composite markers: on the Alternative Agents page the dataframe carries
|
| 1220 |
# an "Agent" column (Claude Code / Codex / Gemini CLI / OpenHands Sub-agents),
|
ui_components.py
CHANGED
|
@@ -43,6 +43,8 @@ from content import (
|
|
| 43 |
api = HfApi()
|
| 44 |
os.makedirs(EXTRACTED_DATA_DIR, exist_ok=True)
|
| 45 |
|
|
|
|
|
|
|
| 46 |
|
| 47 |
def get_company_logo_html(model_name: str) -> str:
|
| 48 |
"""
|
|
@@ -81,12 +83,18 @@ OPENNESS_SVG_MAP = {
|
|
| 81 |
|
| 82 |
def get_svg_as_data_uri(path: str) -> str:
|
| 83 |
"""Reads an SVG file and returns it as a base64-encoded data URI."""
|
|
|
|
|
|
|
|
|
|
| 84 |
try:
|
| 85 |
with open(path, "rb") as svg_file:
|
| 86 |
encoded_svg = base64.b64encode(svg_file.read()).decode("utf-8")
|
| 87 |
-
|
|
|
|
|
|
|
| 88 |
except FileNotFoundError:
|
| 89 |
print(f"Warning: SVG file not found at {path}")
|
|
|
|
| 90 |
return ""
|
| 91 |
|
| 92 |
|
|
|
|
| 43 |
api = HfApi()
|
| 44 |
os.makedirs(EXTRACTED_DATA_DIR, exist_ok=True)
|
| 45 |
|
| 46 |
+
_SVG_DATA_URI_CACHE: dict[str, str] = {}
|
| 47 |
+
|
| 48 |
|
| 49 |
def get_company_logo_html(model_name: str) -> str:
|
| 50 |
"""
|
|
|
|
| 83 |
|
| 84 |
def get_svg_as_data_uri(path: str) -> str:
|
| 85 |
"""Reads an SVG file and returns it as a base64-encoded data URI."""
|
| 86 |
+
if path in _SVG_DATA_URI_CACHE:
|
| 87 |
+
return _SVG_DATA_URI_CACHE[path]
|
| 88 |
+
|
| 89 |
try:
|
| 90 |
with open(path, "rb") as svg_file:
|
| 91 |
encoded_svg = base64.b64encode(svg_file.read()).decode("utf-8")
|
| 92 |
+
uri = f"data:image/svg+xml;base64,{encoded_svg}"
|
| 93 |
+
_SVG_DATA_URI_CACHE[path] = uri
|
| 94 |
+
return uri
|
| 95 |
except FileNotFoundError:
|
| 96 |
print(f"Warning: SVG file not found at {path}")
|
| 97 |
+
_SVG_DATA_URI_CACHE[path] = ""
|
| 98 |
return ""
|
| 99 |
|
| 100 |
|
visualizations.py
CHANGED
|
@@ -108,7 +108,17 @@ def create_accuracy_by_size_chart(df: pd.DataFrame, mark_by: str = None) -> go.F
|
|
| 108 |
open_aliases = [aliases.CANONICAL_OPENNESS_OPEN] + list(
|
| 109 |
aliases.OPENNESS_ALIASES.get(aliases.CANONICAL_OPENNESS_OPEN, [])
|
| 110 |
)
|
| 111 |
-
openness_col =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
|
| 113 |
plot_df = df[
|
| 114 |
(df[param_col].notna()) &
|
|
|
|
| 108 |
open_aliases = [aliases.CANONICAL_OPENNESS_OPEN] + list(
|
| 109 |
aliases.OPENNESS_ALIASES.get(aliases.CANONICAL_OPENNESS_OPEN, [])
|
| 110 |
)
|
| 111 |
+
openness_col = _find_column(df, ['Openness', 'openness'])
|
| 112 |
+
if openness_col is None:
|
| 113 |
+
fig = go.Figure()
|
| 114 |
+
fig.add_annotation(
|
| 115 |
+
text="No openness data available",
|
| 116 |
+
xref="paper", yref="paper",
|
| 117 |
+
x=0.5, y=0.5, showarrow=False,
|
| 118 |
+
font=STANDARD_FONT
|
| 119 |
+
)
|
| 120 |
+
fig.update_layout(**STANDARD_LAYOUT, title="Open Model Accuracy by Size")
|
| 121 |
+
return fig
|
| 122 |
|
| 123 |
plot_df = df[
|
| 124 |
(df[param_col].notna()) &
|