thinking-mode arm: enable_thinking + think-block splitting + think-closing force tail
Browse files- calibration/run.py +18 -6
calibration/run.py
CHANGED
|
@@ -75,6 +75,7 @@ def run_grid(
|
|
| 75 |
strip_schema_legend=False,
|
| 76 |
prose_legend=False,
|
| 77 |
reasoning_note=False,
|
|
|
|
| 78 |
):
|
| 79 |
"""Run every (model, budget) pair through harness.B.launch -- the unmodified
|
| 80 |
extended path, only ``reasoning_budget`` varies -- on the operator's questions."""
|
|
@@ -89,12 +90,16 @@ def run_grid(
|
|
| 89 |
plan = build_plan(models, budgets)
|
| 90 |
for index, (model, budget) in enumerate(plan, start=1):
|
| 91 |
print(f"=== calibration {index}/{len(plan)}: {model} @ {budget} tokens ===", flush=True)
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
harness_kwargs = {}
|
| 99 |
if prose_legend:
|
| 100 |
from harness.B.prompts import PROSE_LEGEND
|
|
@@ -116,6 +121,7 @@ def run_grid(
|
|
| 116 |
question_ids=question_ids,
|
| 117 |
strip_schema_legend=strip_schema_legend or prose_legend,
|
| 118 |
reasoning_note=reasoning_note,
|
|
|
|
| 119 |
**harness_kwargs,
|
| 120 |
)
|
| 121 |
|
|
@@ -164,6 +170,11 @@ def main():
|
|
| 164 |
help="Thinking-with-Spatial-Code step-by-step note prefixed to the "
|
| 165 |
"post-prompt; results land in a '<budget>-reasoning' sibling directory",
|
| 166 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
parser.add_argument("--rebuild", action="store_true")
|
| 168 |
args = parser.parse_args()
|
| 169 |
|
|
@@ -203,6 +214,7 @@ def main():
|
|
| 203 |
strip_schema_legend=args.strip_schema_legend,
|
| 204 |
prose_legend=args.prose_legend,
|
| 205 |
reasoning_note=args.reasoning_note,
|
|
|
|
| 206 |
)
|
| 207 |
except ValueError as exc:
|
| 208 |
parser.error(str(exc))
|
|
|
|
| 75 |
strip_schema_legend=False,
|
| 76 |
prose_legend=False,
|
| 77 |
reasoning_note=False,
|
| 78 |
+
thinking=False,
|
| 79 |
):
|
| 80 |
"""Run every (model, budget) pair through harness.B.launch -- the unmodified
|
| 81 |
extended path, only ``reasoning_budget`` varies -- on the operator's questions."""
|
|
|
|
| 90 |
plan = build_plan(models, budgets)
|
| 91 |
for index, (model, budget) in enumerate(plan, start=1):
|
| 92 |
print(f"=== calibration {index}/{len(plan)}: {model} @ {budget} tokens ===", flush=True)
|
| 93 |
+
variant_parts = [str(budget)]
|
| 94 |
+
if thinking:
|
| 95 |
+
variant_parts.append("thinking")
|
| 96 |
+
if reasoning_note:
|
| 97 |
+
variant_parts.append("reasoning")
|
| 98 |
+
if prose_legend:
|
| 99 |
+
variant_parts.append("prose-legend")
|
| 100 |
+
elif strip_schema_legend:
|
| 101 |
+
variant_parts.append("no-legend")
|
| 102 |
+
variant = "-".join(variant_parts)
|
| 103 |
harness_kwargs = {}
|
| 104 |
if prose_legend:
|
| 105 |
from harness.B.prompts import PROSE_LEGEND
|
|
|
|
| 121 |
question_ids=question_ids,
|
| 122 |
strip_schema_legend=strip_schema_legend or prose_legend,
|
| 123 |
reasoning_note=reasoning_note,
|
| 124 |
+
thinking=thinking,
|
| 125 |
**harness_kwargs,
|
| 126 |
)
|
| 127 |
|
|
|
|
| 170 |
help="Thinking-with-Spatial-Code step-by-step note prefixed to the "
|
| 171 |
"post-prompt; results land in a '<budget>-reasoning' sibling directory",
|
| 172 |
)
|
| 173 |
+
parser.add_argument(
|
| 174 |
+
"--thinking", action="store_true",
|
| 175 |
+
help="enable native thinking mode (Qwen only); results land in a "
|
| 176 |
+
"'<budget>-thinking[...]' sibling directory",
|
| 177 |
+
)
|
| 178 |
parser.add_argument("--rebuild", action="store_true")
|
| 179 |
args = parser.parse_args()
|
| 180 |
|
|
|
|
| 214 |
strip_schema_legend=args.strip_schema_legend,
|
| 215 |
prose_legend=args.prose_legend,
|
| 216 |
reasoning_note=args.reasoning_note,
|
| 217 |
+
thinking=args.thinking,
|
| 218 |
)
|
| 219 |
except ValueError as exc:
|
| 220 |
parser.error(str(exc))
|