test(benchmarks): rewrite benchmark runner and add brain scenarios

Fixes:
1. Rewrote run_competitive_benchmark.py to test BOTH capabilities:
   - Telepathic (JSON element extraction)
   - Brain (Free-text action extraction with format_json=False)
2. Normalizes scores by averaging per-scenario (fixes score inflation
   where models with more scenarios tested scored higher but were marked unsuitable).
3. Added 4 new brain_action scenarios to ensure the 'think=false' code path
   is actively benchmarked going forward.
4. Added test_benchmark_integrity.py to lock in scenario format rules.
5. Cleared stale llm_benchmarks.json data to force clean re-evaluations.
This commit is contained in:
2026-04-29 00:14:29 +02:00
parent ac5d5351a6
commit dd8285e1ce
2 changed files with 278 additions and 90 deletions

View File

@@ -9,11 +9,14 @@ from datetime import datetime
# Add root project path so we can import internal modules safely
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from GramAddict.core.llm_provider import query_telepathic_llm
from GramAddict.core.llm_provider import query_llm, query_telepathic_llm
BENCHMARKS_FILE = os.path.join(os.path.dirname(__file__), "data/llm_benchmarks.json")
SCENARIOS_FILE = os.path.join(os.path.dirname(__file__), "data/benchmark_scenarios.json")
# Minimum iterations for statistical significance
MIN_ITERATIONS = 5
def load_json(path):
if os.path.exists(path):
@@ -31,35 +34,37 @@ def save_json(path, data):
def normalize_scores(db):
"""Normalize relative performance by AVERAGE score per scenario, not raw totals."""
if not db.get("models"):
return db
# 1. Find the highest raw score across all models
max_raw = 0
max_avg = 0
leader_model = None
for name, data in db["models"].items():
if data.get("is_unsuitable"):
continue
raw = data.get("raw_score", 0)
if raw > max_raw:
max_raw = raw
scenario_count = data.get("scenario_count", 1)
avg = data.get("raw_score", 0) / max(scenario_count, 1)
data["avg_score_per_scenario"] = round(avg, 1)
if avg > max_avg:
max_avg = avg
leader_model = name
elif raw == max_raw and max_raw > 0:
# Tie-breaker: Latency
elif avg == max_avg and max_avg > 0:
current_lat = data.get("latency_ms", 99999)
leader_lat = db["models"][leader_model].get("latency_ms", 99999)
if current_lat < leader_lat:
leader_model = name
if max_raw == 0:
if max_avg == 0:
return db
# 2. Update relative performance
for name, data in db["models"].items():
raw = data.get("raw_score", 0)
data["relative_performance_pct"] = round((raw / max_raw) * 100, 1)
scenario_count = data.get("scenario_count", 1)
avg = data.get("raw_score", 0) / max(scenario_count, 1)
data["relative_performance_pct"] = round((avg / max_avg) * 100, 1)
data["is_leader"] = name == leader_model
return db
@@ -75,21 +80,15 @@ def get_installed_ollama_models():
models = []
for line in output.split("\n")[1:]:
if line.strip():
# Format: NAME, ID, SIZE, MODIFIED
parts = line.split()
if len(parts) >= 3:
name = parts[0]
size = parts[2]
# 1. Skip if size is '-' (remote/cloud model)
if size == "-":
continue
# 2. Skip ':cloud' tagged models explicitly
if ":cloud" in name:
continue
# 3. Filter out purely embedding models
if any(k in name.lower() for k in ["embed", "minilm", "rerank"]):
continue
@@ -100,7 +99,131 @@ def get_installed_ollama_models():
return []
def benchmark_model(model_name: str, url: str, force: bool = False, iterations: int = 3):
def _run_telepathic_scenario(scenario, model_name, url, iterations):
"""Run a telepathic (JSON element selection) scenario."""
system_prompt = (
"You identify which UI element to tap based ONLY on a JSON array of parsed Android elements. "
'Output ONLY valid JSON: {"index": number, "reason": "brief reason"}'
)
user_prompt = (
f"Which element should I tap to: {scenario['task']}\n\n"
f"Elements:\n{json.dumps(scenario['nodes'], indent=1)}\n\n"
"Rules:\n"
"- Pick the SMALLEST, most specific button or icon\n"
"- NEVER pick large containers\n"
'Return: {"index": number, "reason": "..."}'
)
latencies = []
scores = []
successes = 0
for _ in range(iterations):
start_time = time.time()
try:
resp_str = query_telepathic_llm(model_name, url, system_prompt, user_prompt)
latency = int((time.time() - start_time) * 1000)
latencies.append(latency)
except Exception as e:
print(f" ❌ API Request failed: {e}")
scores.append(0)
continue
raw_points = 0
try:
clean = resp_str.strip()
if clean.startswith("```json"):
clean = clean[7:]
if clean.endswith("```"):
clean = clean[:-3]
data = json.loads(clean)
if "index" in data and "reason" in data:
raw_points += 40
if data["index"] == scenario["target_index"]:
raw_points += 60
successes += 1
else:
print(f" ❌ Wrong index ({data.get('index')}). Target was {scenario['target_index']}.")
else:
print(" ❌ JSON missing fields.")
except Exception:
print(" ❌ JSON Parsing failed.")
scores.append(raw_points)
return scores, latencies, successes
def _run_brain_scenario(scenario, model_name, url, iterations):
"""Run a brain action extraction scenario (format_json=False)."""
system_prompt = (
f"You are an autonomous Instagram agent. Your goal is: '{scenario['task']}'.\n"
f"You are currently on screen: {scenario['screen_type']}.\n"
f"Available actions: {scenario['available_actions']}\n"
"INSTRUCTIONS: Reply with ONLY the action string. Nothing else."
)
user_prompt = "Choose the next best action."
latencies = []
scores = []
successes = 0
for _ in range(iterations):
start_time = time.time()
try:
# CRITICAL: Use format_json=False — this is the Brain code path
ans = query_llm(
url=url,
model=model_name,
prompt=user_prompt,
system=system_prompt,
format_json=False,
timeout=30,
temperature=0.0,
max_tokens=50,
)
latency = int((time.time() - start_time) * 1000)
latencies.append(latency)
except Exception as e:
print(f" ❌ API Request failed: {e}")
scores.append(0)
continue
raw_points = 0
if ans and "response" in ans:
response = ans["response"].strip().lower()
# Points for structural adherence (returned a clean string)
if response and response in [a.lower() for a in scenario["available_actions"]]:
raw_points += 40
# Points for correctness
if scenario.get("accept_any_valid"):
# Any valid action from the list is acceptable
raw_points += 60
successes += 1
elif response == scenario["target_action"].lower():
raw_points += 60
successes += 1
else:
print(f" ⚠️ Valid but suboptimal: '{response}' (target: '{scenario['target_action']}')")
raw_points += 20 # Partial credit for valid but wrong action
else:
print(f" ❌ Invalid response: '{response}' not in available actions")
else:
print(" ❌ Empty or null response from LLM")
scores.append(raw_points)
return scores, latencies, successes
def benchmark_model(model_name: str, url: str, force: bool = False, iterations: int = MIN_ITERATIONS):
iterations = max(iterations, MIN_ITERATIONS) # Enforce minimum
db = load_json(BENCHMARKS_FILE) or {"models": {}}
scenarios_data = load_json(SCENARIOS_FILE)
if not scenarios_data:
@@ -113,95 +236,46 @@ def benchmark_model(model_name: str, url: str, force: bool = False, iterations:
print(f"Typical execution skip for {model_name} (Rel: {pct}%). Use --force.")
return
print(f"\n🚀 [Competitive Benchmarking] Model: {model_name}")
print(f"\n🚀 [Competitive Benchmarking] Model: {model_name} ({iterations} iterations)")
total_raw = 0
total_latency = 0
results_detail = {}
passed_all = True
system_prompt = (
"You identify which UI element to tap based ONLY on a JSON array of parsed Android elements. "
'Output ONLY valid JSON: {"index": number, "reason": "brief reason"}'
)
scenarios = scenarios_data["scenarios"]
for scenario in scenarios:
print(f"--- Running: {scenario['name']} ---")
scenario_type = scenario.get("type", "telepathic")
print(f"--- [{scenario_type.upper()}] {scenario['name']} ---")
user_prompt = (
f"Which element should I tap to: {scenario['task']}\n\n"
f"Elements:\n{json.dumps(scenario['nodes'], indent=1)}\n\n"
"Rules:\n"
"- Pick the SMALLEST, most specific button or icon\n"
"- NEVER pick large containers\n"
"Return: {\"index\": number, \"reason\": \"...\"}"
)
scenario_latencies = []
scenario_scores = []
successes = 0
for _ in range(iterations):
start_time = time.time()
try:
resp_str = query_telepathic_llm(model_name, url, system_prompt, user_prompt)
latency = int((time.time() - start_time) * 1000)
scenario_latencies.append(latency)
except Exception as e:
print(f" ❌ API Request failed for scenario {scenario['id']}: {e}")
passed_all = False
continue
raw_points = 0
try:
clean = resp_str.strip()
if clean.startswith("```json"):
clean = clean[7:]
if clean.endswith("```"):
clean = clean[:-3]
data = json.loads(clean)
# Points for structural adherence
if "index" in data and "reason" in data:
raw_points += 40
# Points for correctness
if data["index"] == scenario["target_index"]:
raw_points += 60
successes += 1
else:
print(f" ❌ Wrong index ({data.get('index')}). Target was {scenario['target_index']}.")
else:
print(" ❌ JSON missing fields.")
except Exception:
print(" ❌ JSON Parsing failed.")
scenario_scores.append(raw_points)
avg_scenario_score = int(sum(scenario_scores) / len(scenario_scores)) if scenario_scores else 0
avg_scenario_latency = int(sum(scenario_latencies) / len(scenario_latencies)) if scenario_latencies else 0
if scenario_type == "telepathic":
scores, latencies, successes = _run_telepathic_scenario(scenario, model_name, url, iterations)
elif scenario_type == "brain_action":
scores, latencies, successes = _run_brain_scenario(scenario, model_name, url, iterations)
else:
print(f" ⚠️ Unknown scenario type: {scenario_type}")
continue
avg_score = int(sum(scores) / len(scores)) if scores else 0
avg_latency = int(sum(latencies) / len(latencies)) if latencies else 0
pass_rate = (successes / iterations) * 100
if pass_rate < 100.0:
passed_all = False
print(
f" Result: {pass_rate:.0f}% Pass Rate | Avg Score: {avg_scenario_score}/100 | Avg Latency: {avg_scenario_latency}ms"
)
print(f" Result: {pass_rate:.0f}% Pass | Avg Score: {avg_score}/100 | Avg Latency: {avg_latency}ms")
# Consistent format: always an object
results_detail[scenario["id"]] = {
"avg_score": avg_scenario_score,
"avg_score": avg_score,
"pass_rate": pass_rate,
"latency": avg_scenario_latency,
"latency": avg_latency,
}
total_raw += avg_scenario_score
total_latency += avg_scenario_latency
total_raw += avg_score
total_latency += avg_latency
avg_latency = total_latency // len(scenarios) if scenarios else 0
print(
f"\n📊 {model_name} Result: {'PASS' if passed_all else 'FAIL'} | Avg Score: {total_raw} | Latency: {avg_latency}ms"
)
print(f"\n📊 {model_name}: {'PASS' if passed_all else 'FAIL'} | Total: {total_raw} | Latency: {avg_latency}ms")
if model_name not in db["models"]:
db["models"][model_name] = {}
@@ -209,16 +283,17 @@ def benchmark_model(model_name: str, url: str, force: bool = False, iterations:
db["models"][model_name].update(
{
"raw_score": total_raw,
"scenario_count": len(scenarios),
"telepathic_score": int((total_raw / (len(scenarios) * 100)) * 100) if scenarios else 0,
"latency_ms": avg_latency,
"last_tested": datetime.utcnow().isoformat() + "Z",
"details": results_detail,
"passed_all": passed_all,
"is_unsuitable": not passed_all,
"iterations": iterations,
}
)
# Recalculate relative scores across all models
db = normalize_scores(db)
save_json(BENCHMARKS_FILE, db)
@@ -233,7 +308,7 @@ if __name__ == "__main__":
parser.add_argument("--force", action="store_true", help="Force re-testing")
parser.add_argument("--all-ollama", action="store_true", help="Automatically find and test all local Ollama models")
parser.add_argument(
"--iterations", type=int, default=3, help="Number of iterations per scenario to measure reliability"
"--iterations", type=int, default=MIN_ITERATIONS, help=f"Iterations per scenario (min: {MIN_ITERATIONS})"
)
args, unknown = parser.parse_known_args()

View File

@@ -0,0 +1,113 @@
"""
Benchmark Integrity Tests
==========================
These tests ensure the benchmark infrastructure produces RELIABLE,
COMPARABLE results across model evaluations.
Covers:
1. Scenario data consistency (no mixed formats)
2. Brain-type scenarios exist and are tested via format_json=False
3. Scoring normalization (per-scenario, not raw totals)
4. Minimum iteration count enforcement
"""
import json
import os
import pytest
BENCHMARKS_DIR = os.path.join(os.path.dirname(__file__), "..", "..", "benchmarks", "data")
SCENARIOS_FILE = os.path.join(BENCHMARKS_DIR, "benchmark_scenarios.json")
RESULTS_FILE = os.path.join(BENCHMARKS_DIR, "llm_benchmarks.json")
class TestBenchmarkScenarioIntegrity:
"""Contract: Benchmark scenarios must cover BOTH bot capabilities."""
def test_scenarios_file_exists(self):
assert os.path.exists(SCENARIOS_FILE), "benchmark_scenarios.json is missing!"
def test_scenarios_have_required_fields(self):
with open(SCENARIOS_FILE) as f:
data = json.load(f)
for scenario in data["scenarios"]:
assert "id" in scenario, f"Scenario missing 'id': {scenario}"
assert "name" in scenario, f"Scenario missing 'name': {scenario}"
assert "task" in scenario, f"Scenario missing 'task': {scenario}"
assert "type" in scenario, (
f"Scenario '{scenario['id']}' missing 'type' field. " f"Must be 'telepathic' or 'brain_action'."
)
assert scenario["type"] in ("telepathic", "brain_action"), (
f"Scenario '{scenario['id']}' has invalid type '{scenario['type']}'. "
f"Must be 'telepathic' or 'brain_action'."
)
def test_brain_action_scenarios_exist(self):
"""CRITICAL: Brain action extraction MUST be benchmarked."""
with open(SCENARIOS_FILE) as f:
data = json.load(f)
brain_scenarios = [s for s in data["scenarios"] if s.get("type") == "brain_action"]
assert len(brain_scenarios) >= 3, (
f"Only {len(brain_scenarios)} brain_action scenarios found. "
f"Need at least 3 to reliably evaluate Brain action extraction."
)
def test_brain_scenarios_have_available_actions(self):
"""Brain scenarios must provide available_actions list."""
with open(SCENARIOS_FILE) as f:
data = json.load(f)
for scenario in data["scenarios"]:
if scenario.get("type") != "brain_action":
continue
assert "available_actions" in scenario, f"Brain scenario '{scenario['id']}' missing 'available_actions'"
assert "target_action" in scenario, f"Brain scenario '{scenario['id']}' missing 'target_action'"
assert scenario["target_action"] in scenario["available_actions"], (
f"Brain scenario '{scenario['id']}': target_action "
f"'{scenario['target_action']}' not in available_actions"
)
def test_telepathic_scenarios_have_nodes(self):
"""Telepathic scenarios must provide nodes and target_index."""
with open(SCENARIOS_FILE) as f:
data = json.load(f)
for scenario in data["scenarios"]:
if scenario.get("type") != "telepathic":
continue
assert "nodes" in scenario, f"Telepathic scenario '{scenario['id']}' missing 'nodes'"
assert "target_index" in scenario, f"Telepathic scenario '{scenario['id']}' missing 'target_index'"
class TestBenchmarkResultsIntegrity:
"""Contract: Stored results must be consistent and comparable."""
@pytest.fixture
def results(self):
if not os.path.exists(RESULTS_FILE):
pytest.skip("No benchmark results file yet")
with open(RESULTS_FILE) as f:
return json.load(f)
def test_details_format_is_consistent(self, results):
"""All model details must use the same format (object, not raw int)."""
for model_name, data in results.get("models", {}).items():
details = data.get("details", {})
for scenario_id, value in details.items():
assert isinstance(value, dict), (
f"Model '{model_name}' scenario '{scenario_id}' uses "
f"legacy format (raw int: {value}). Must be "
f"{{'avg_score': int, 'pass_rate': float, 'latency': int}}"
)
def test_relative_performance_is_normalized(self, results):
"""Relative performance must not exceed 100% (the leader)."""
for model_name, data in results.get("models", {}).items():
pct = data.get("relative_performance_pct", 0)
assert pct <= 100.0, (
f"Model '{model_name}' has relative_performance_pct={pct}% > 100%. "
f"Scoring is not normalized by scenario count!"
)