Datasets:
File size: 10,985 Bytes
f79f2ee | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 | import argparse
import json
import os
import re
from openai import OpenAI
import anthropic
from dotenv import load_dotenv
load_dotenv()
OpenAI_API_KEY = os.getenv("OpenAI_API_KEY")
Anthropic_API_KEY = os.getenv("Anthropic_API_KEY")
JUDGE_CORRECT_PROMPT = (
"You are evaluating whether a vision model correctly identified an occluded object.\n\n"
"Ground truth: \"{ground_truth}\"\n\n"
"Model response: \"{response}\"\n\n"
"Does the model's response correctly identify the same object as the ground truth? "
"Allow for reasonable variation in phrasing, specificity, or synonyms "
"(e.g., 'barber pole' matches 'barber's pole', 'wing mirror' matches 'side mirror', "
"'cat' matches 'pet' if the ground truth is 'Reasonable Pet'). "
"The response should identify the core object — partial credit guesses that get the "
"category right but miss specifics should be marked correct only if the category-level "
"answer is what the ground truth describes. "
"Exception: if the ground truth explicitly refers to a reflection or shadow of an object "
"(e.g., 'pen shadow', 'reflection of a car'), the response must identify both the shadow/reflection "
"nature AND the underlying object to be marked correct — identifying only the object (e.g., 'pen') "
"is not sufficient. "
"Answer only 'yes' or 'no'."
)
def judge_correct(ground_truth, response, judge_model, clients):
prompt = JUDGE_CORRECT_PROMPT.format(
ground_truth=ground_truth.strip(),
response=response.strip()
)
if judge_model.startswith("claude"):
verdict = clients["anthropic"].messages.create(
model=judge_model,
temperature=0,
max_tokens=10,
messages=[{"role": "user", "content": prompt}]
).content[0].text.strip().lower()
else:
verdict = clients["openai"].chat.completions.create(
model=judge_model,
temperature=0,
messages=[{"role": "user", "content": prompt}]
).choices[0].message.content.strip().lower()
return verdict.startswith("yes")
def get_effective_response(entry):
"""Pick the best response from an occluded-eval entry."""
for key in ("mask_followup_response", "followup_response", "response"):
val = entry.get(key)
if val:
return val, key
return "", "response"
def discover_model_stems(results_dir):
"""Return all unique model stems found in the results directory."""
stems = set()
for fname in os.listdir(results_dir):
if not fname.endswith(".json") or fname.endswith("_judged_results.json"):
continue
for suffix in ("_vetted_results.json", "_transparent_results.json", "_results.json"):
if fname.endswith(suffix):
stems.add(fname[: -len(suffix)])
break
return sorted(stems)
def load_results(results_dir, stem):
"""Load occluded results: prefer vetted, fall back to plain results."""
vetted_path = os.path.join(results_dir, f"{stem}_vetted_results.json")
plain_path = os.path.join(results_dir, f"{stem}_results.json")
if os.path.exists(vetted_path):
with open(vetted_path) as f:
return json.load(f), "vetted"
elif os.path.exists(plain_path):
with open(plain_path) as f:
return json.load(f), "plain"
return None, None
def load_transparent_results(results_dir, stem):
path = os.path.join(results_dir, f"{stem}_transparent_results.json")
if os.path.exists(path):
with open(path) as f:
return json.load(f)
return None
def run_judger(args):
clients = {
"openai": OpenAI(api_key=OpenAI_API_KEY),
"anthropic": anthropic.Anthropic(api_key=Anthropic_API_KEY),
}
os.makedirs(args.output_dir, exist_ok=True)
judge_models = args.judge_models
stems = discover_model_stems(args.results_dir)
if not stems:
print("No result files found.")
return
print(f"Found models: {', '.join(stems)}\n")
print(f"Judge models: {', '.join(judge_models)}\n")
for stem in stems:
occluded_results, source_type = load_results(args.results_dir, stem)
transparent_results = load_transparent_results(args.results_dir, stem)
has_occluded = occluded_results is not None
has_transparent = transparent_results is not None
if not has_occluded and not has_transparent:
print(f"[{stem}] No usable result files found, skipping.")
continue
output_path = os.path.join(args.output_dir, f"{stem}_judged_results.json")
# Load existing judged output for resume support
if os.path.exists(output_path):
with open(output_path) as f:
judged = json.load(f)
judged_by_id = {r["id"]: r for r in judged}
print(f"[{stem}] Resuming: {len(judged_by_id)} already judged.")
else:
judged_by_id = {}
# Build lookup for transparent results
transparent_by_id = {}
if has_transparent:
for entry in transparent_results:
transparent_by_id[entry["id"]] = entry
# Collect all IDs to judge
all_ids = set()
occluded_by_id = {}
if has_occluded:
for entry in occluded_results:
occluded_by_id[entry["id"]] = entry
all_ids.add(entry["id"])
all_ids.update(transparent_by_id.keys())
updated = False
for entry_id in sorted(all_ids, key=lambda x: (0, int(x)) if x.isdigit() else (1, x)):
existing = judged_by_id.get(entry_id, {"id": entry_id})
# Ensure verdict dicts exist
if not isinstance(existing.get("occluded_correct"), dict):
existing["occluded_correct"] = {}
if not isinstance(existing.get("transparent_correct"), dict):
existing["transparent_correct"] = {}
# Always sync tool_calls from raw results
if entry_id in occluded_by_id and "tool_calls" in occluded_by_id[entry_id]:
existing["tool_calls"] = occluded_by_id[entry_id]["tool_calls"]
# Determine which judge models still need to run
pending_occ = entry_id in occluded_by_id and [
jm for jm in judge_models if jm not in existing["occluded_correct"]
]
pending_trans = entry_id in transparent_by_id and [
jm for jm in judge_models if jm not in existing["transparent_correct"]
]
if not pending_occ and not pending_trans:
continue
print(f"[{stem}] Judging {entry_id}...")
# --- Occluded eval ---
if pending_occ:
occ = occluded_by_id[entry_id]
ground_truth = occ.get("ground_truth", "").strip()
effective_response, response_key = get_effective_response(occ)
existing["ground_truth"] = ground_truth
existing["occluded_source"] = source_type
existing["occluded_response_used"] = response_key
existing["occluded_response"] = effective_response
for jm in pending_occ:
if ground_truth and effective_response:
correct = judge_correct(ground_truth, effective_response, jm, clients)
existing["occluded_correct"][jm] = correct
print(f" occluded ({response_key}) [{jm}]: {'CORRECT' if correct else 'WRONG'}")
else:
existing["occluded_correct"][jm] = None
print(f" occluded [{jm}]: skipped (missing ground truth or response)")
# --- Transparent eval ---
if pending_trans:
trans = transparent_by_id[entry_id]
ground_truth = trans.get("ground_truth", "").strip()
trans_response = trans.get("response", "")
if "ground_truth" not in existing:
existing["ground_truth"] = ground_truth
existing["transparent_response"] = trans_response
for jm in pending_trans:
if ground_truth and trans_response:
correct = judge_correct(ground_truth, trans_response, jm, clients)
existing["transparent_correct"][jm] = correct
print(f" transparent [{jm}]: {'CORRECT' if correct else 'WRONG'}")
else:
existing["transparent_correct"][jm] = None
print(f" transparent [{jm}]: skipped (missing ground truth or response)")
judged_by_id[entry_id] = existing
updated = True
with open(output_path, "w") as f:
json.dump(list(judged_by_id.values()), f, indent=2)
if not updated and judged_by_id:
with open(output_path, "w") as f:
json.dump(list(judged_by_id.values()), f, indent=2)
# Summary per judge model
all_judged = list(judged_by_id.values())
print(f"\n[{stem}] Summary:")
for jm in judge_models:
occ_judged = [r for r in all_judged if isinstance(r.get("occluded_correct"), dict) and r["occluded_correct"].get(jm) is not None]
trans_judged = [r for r in all_judged if isinstance(r.get("transparent_correct"), dict) and r["transparent_correct"].get(jm) is not None]
occ_correct = sum(1 for r in occ_judged if r["occluded_correct"][jm])
trans_correct = sum(1 for r in trans_judged if r["transparent_correct"][jm])
print(f" [{jm}]")
if occ_judged:
print(f" Occluded ({source_type}): {occ_correct}/{len(occ_judged)} correct "
f"({100 * occ_correct / len(occ_judged):.1f}%)")
if trans_judged:
print(f" Transparent: {trans_correct}/{len(trans_judged)} correct "
f"({100 * trans_correct / len(trans_judged):.1f}%)")
print(f" Output: {output_path}\n")
def parse_args():
parser = argparse.ArgumentParser(
description="Judge model responses against ground truth for occluded and transparent evals."
)
parser.add_argument("--results_dir", type=str, default="data/eval_results",
help="Directory containing *_results.json / *_vetted_results.json / *_transparent_results.json files.")
parser.add_argument("--output_dir", type=str, default="data/eval_results",
help="Directory to write *_judged_results.json files.")
parser.add_argument("--judge_models", type=str, nargs="+", default=["gpt-5.4"],
help="One or more OpenAI models to use for judging.")
return parser.parse_args()
if __name__ == "__main__":
args = parse_args()
run_judger(args)
|