agent-cost-optimizer / training /train_router_v2.py
narcolepticchicken's picture
Upload training/train_router_v2.py with huggingface_hub
2912ad5 verified
#!/usr/bin/env python3
"""Trained Router v2: Safety-first CARROT with tuned mu + safety floors.
Key insight from v1:
- Per-tier P(success) classifiers work well individually
- CARROT routing with mu=0.6 beats heuristic on both quality and cost
- But success rate drops because CARROT routes cheap for hard tasks
Solution: Add SAFETY FLOORS per task type:
- legal_regulated: never below tier 4
- coding/research with legal kw: never below tier 3
- Use P(success) > threshold as gate, fallback to difficulty-based tier
- When confidence is low, default to tier 3 (medium)
"""
import json, os, sys, random, pickle, uuid
import numpy as np
from datetime import datetime
from collections import defaultdict
TASK_TYPES = ["quick_answer","coding","research","document_drafting",
"legal_regulated","tool_heavy","retrieval_heavy",
"long_horizon","unknown_ambiguous"]
TT2IDX = {t:i for i,t in enumerate(TASK_TYPES)}
CODE_KW = ["python","javascript","code","function","bug","debug","refactor",
"implement","test","compile","runtime","class","module","async","thread"]
LEGAL_KW = ["contract","legal","compliance","gdpr","privacy","policy","regulatory","liability"]
RESEARCH_KW = ["research","find sources","literature","investigate","compare","analyze","survey"]
TOOL_KW = ["search","fetch","retrieve","query","api","database","scrape","aggregate"]
LONG_KW = ["plan","project","roadmap","orchestrate","multi-step","migrate","pipeline","deploy"]
MATH_KW = ["calculate","compute","solve","equation","formula","optimize","probability"]
TIER_STR = {1:0.35,2:0.55,3:0.80,4:0.93,5:0.97}
TIER_COST = {1:0.05,2:0.15,3:0.75,4:1.0,5:1.5}
TASK_TEMPLATES = {
"quick_answer":["What is the capital of France?","Explain quantum computing briefly.",
"What is 237*452?","Define photosynthesis.","Who wrote Hamlet?",
"What is the speed of light?","List the primary colors.","What is GDP?"],
"coding":["Write a Python function to reverse a linked list.",
"Fix the bug in this React component.","Refactor auth module to JWT.",
"Implement LRU cache in Go.","Debug segfault in C++ thread pool.",
"Add unit tests for the payment module.","Optimize this SQL query.",
"Create a REST API for user management.","Implement binary search in Rust."],
"research":["Research latest transformer advances.",
"Find sources comparing LoRA and full FT.",
"Investigate data center climate impact.",
"Survey privacy-preserving ML techniques.",
"Compare reinforcement learning algorithms for robotics."],
"document_drafting":["Draft project proposal for ML pipeline.",
"Write email to team about deployment.","Create technical report on performance."],
"legal_regulated":["Review this contract for liability clauses.",
"Check GDPR compliance for data pipeline.","Draft privacy policy section.",
"Verify regulatory compliance for medical device software."],
"tool_heavy":["Search open issues and create summary.",
"Fetch API docs and generate client code.","Query Q3 sales and produce chart."],
"retrieval_heavy":["Answer based on 50-page document.",
"Find all payment processing mentions.","Retrieve relevant cases for legal query."],
"long_horizon":["Plan 3-month roadmap.","Orchestrate multi-region deployment.",
"Redesign data architecture end-to-end.","Migrate monolith to microservices."],
"unknown_ambiguous":["Help me with this thing.",
"I need something about the server.","Can you look into that issue?"],
}
# Safety floors per task type
TASK_FLOOR = {
"legal_regulated": 4,
"long_horizon": 3,
"research": 3,
"coding": 3,
"unknown_ambiguous": 3,
"quick_answer": 1,
"document_drafting": 2,
"tool_heavy": 2,
"retrieval_heavy": 2,
}
def tsp(tier, diff):
return TIER_STR[tier] ** (diff * 0.6)
def extract_features(request, task_type, difficulty=3):
r = request.lower()
f = {
"req_len": len(request),
"num_words": len(request.split()),
"has_code": int(any(k in r for k in CODE_KW)),
"n_code": sum(1 for k in CODE_KW if k in r),
"has_legal": int(any(k in r for k in LEGAL_KW)),
"n_legal": sum(1 for k in LEGAL_KW if k in r),
"has_research": int(any(k in r for k in RESEARCH_KW)),
"n_research": sum(1 for k in RESEARCH_KW if k in r),
"has_tool": int(any(k in r for k in TOOL_KW)),
"n_tool": sum(1 for k in TOOL_KW if k in r),
"has_long": int(any(k in r for k in LONG_KW)),
"has_math": int(any(k in r for k in MATH_KW)),
"tt_idx": TT2IDX.get(task_type, 8),
"difficulty": difficulty,
}
for tt in TASK_TYPES:
f[f"tt_{tt}"] = int(task_type == tt)
return f
def gen_trace(idx, rng):
tt = rng.choice(list(TASK_TEMPLATES.keys()))
diff = {"quick_answer":1,"document_drafting":2,"tool_heavy":2,"retrieval_heavy":2,
"research":3,"coding":3,"unknown_ambiguous":3,"long_horizon":4,"legal_regulated":5}[tt]
tier_out = {}
for t in range(1,6):
tier_out[t] = rng.random() < tsp(t, diff)
opt = 5
for t in range(1,6):
if tier_out[t]:
opt = t
break
if diff <= 2:
actual = rng.choices([1,2,3,4,5],weights=[3,4,2,1,0.5])[0]
elif diff == 3:
actual = rng.choices([1,2,3,4,5],weights=[1,2,4,2,1])[0]
elif diff == 4:
actual = rng.choices([1,2,3,4,5],weights=[0.5,1,2,4,2])[0]
else:
actual = rng.choices([1,2,3,4,5],weights=[0.2,0.5,1,3,4])[0]
outcome = "success" if tier_out[actual] else "failure"
req = rng.choice(TASK_TEMPLATES[tt])
feats = extract_features(req, tt, diff)
return {"feats":feats,"opt":opt,"actual":actual,"outcome":outcome,
"tier_out":tier_out,"tt":tt,"diff":diff,"req":req}
print("="*80)
print("AGENT COST OPTIMIZER - TRAINED ROUTER v2 (Safety-First CARROT)")
print("="*80)
print("\n[1] Generating 50K training traces...")
rng = random.Random(42)
traces = [gen_trace(i, rng) for i in range(50000)]
print(f" Generated {len(traces)} traces")
# Feature matrix
FEAT_KEYS = sorted(traces[0]["feats"].keys())
NUM_FEATURES = len(FEAT_KEYS)
def f2v(feats):
return np.array([float(feats.get(k, 0.0)) for k in FEAT_KEYS], dtype=np.float32)
X_all = np.array([f2v(t["feats"]) for t in traces])
y_opt = np.array([t["opt"] for t in traces])
# Per-tier labels
per_tier_labels = {}
for tier in range(1, 6):
per_tier_labels[tier] = np.array([1 if t["tier_out"].get(tier, False) else 0 for t in traces])
# Train/test split
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, f1_score
X_train, X_test, idx_train, idx_test = train_test_split(
X_all, range(len(traces)), test_size=0.2, random_state=42, stratify=y_opt
)
print(f" Train: {len(X_train)}, Test: {len(X_test)}")
# ─── Train Per-Tier XGBoost Classifiers ────────────────────────────
print("\n[2] Training per-tier P(success) XGBoost classifiers...")
import xgboost as xgb
tier_clfs = {}
for tier in range(1, 6):
y_tr = per_tier_labels[tier][idx_train]
y_te = per_tier_labels[tier][idx_test]
# Compute scale_pos_weight for imbalanced classes
neg = (y_tr == 0).sum()
pos = (y_tr == 1).sum()
spw = neg / max(pos, 1)
clf = xgb.XGBClassifier(
n_estimators=150, max_depth=5, learning_rate=0.1,
subsample=0.8, colsample_bytree=0.8,
scale_pos_weight=min(spw, 5.0),
objective="binary:logistic", eval_metric="logloss",
random_state=42, verbosity=0,
)
clf.fit(X_train, y_tr)
y_pred = clf.predict(X_test)
acc = accuracy_score(y_te, y_pred)
f1 = f1_score(y_te, y_pred, zero_division=0)
tier_clfs[tier] = clf
print(f" Tier {tier}: acc={acc:.3f}, f1={f1:.3f}, spw={spw:.2f}")
# ─── Safety-First CARROT Router ─────────────────────────────────────
print("\n[3] Building safety-first CARROT router...")
def route_safe_carrot(features_vec, tier_clfs, task_type, mu=0.7,
success_threshold=0.5, safety_floor=None):
"""Route with safety floors.
1. Compute P(success|tier) for each tier
2. Apply safety floor per task type
3. Pick cheapest tier where P(success) > threshold
4. If none meets threshold, escalate to next tier
"""
if features_vec.ndim == 1:
features_vec = features_vec.reshape(1, -1)
floor = safety_floor or TASK_FLOOR.get(task_type, 2)
# Get per-tier success probabilities
p_success = {}
for tier in range(1, 6):
p_success[tier] = tier_clfs[tier].predict_proba(features_vec)[0, 1]
# Strategy: Find cheapest tier at or above floor where P(success) > threshold
for tier in range(floor, 6):
if p_success[tier] >= success_threshold:
return tier, p_success
# Fallback: if no tier meets threshold at floor, try escalating
for tier in range(floor + 1, 6):
if p_success[tier] >= success_threshold * 0.8: # relaxed threshold
return tier, p_success
# Last resort: use CARROT scoring at floor
best_tier = floor
best_score = float("inf")
for tier in range(floor, 6):
cost_norm = TIER_COST[tier] / TIER_COST[5]
score = mu * (1.0 - p_success[tier]) + (1.0 - mu) * cost_norm
if score < best_score:
best_score = score
best_tier = tier
return best_tier, p_success
# ─── Evaluate ────────────────────────────────────────────────────────
print("\n[4] Evaluating all routers on test set...")
n_test = len(idx_test)
results = {}
# Helper: evaluate a router function
def eval_router(name, route_fn):
succ = 0; cost = 0.0; unsafe = 0; false_done = 0
tier_dist = defaultdict(int)
for i in idx_test:
t = traces[i]
x = f2v(t["feats"]).reshape(1, -1)
pred, _ = route_fn(x, t)
tier_dist[pred] += 1
if t["tier_out"].get(pred, False):
succ += 1
else:
if pred < t["opt"]:
unsafe += 1
if pred >= t["opt"] and not t["tier_out"].get(pred, False):
false_done += 1
cost += TIER_COST[pred]
results[name] = {
"success": succ/n_test, "avg_cost": cost/n_test,
"unsafe_rate": unsafe/n_test, "false_done": false_done/n_test,
"tier_dist": dict(tier_dist),
}
# 1. Always frontier
eval_router("always_frontier", lambda x, t: (4, {}))
# 2. Always cheapest
eval_router("always_cheap", lambda x, t: (1, {}))
# 3. Heuristic (difficulty + 1)
eval_router("heuristic_diff+1", lambda x, t: (min(t["diff"]+1, 5), {}))
# 4. Heuristic (task floor only)
eval_router("heuristic_floor", lambda x, t: (TASK_FLOOR.get(t["tt"], 3), {}))
# 5. CARROT v1 (no safety floors, mu=0.6)
def carrot_v1(x, t):
ps = {tier: tier_clfs[tier].predict_proba(x)[0,1] for tier in range(1,6)}
best = 3; best_s = float("inf")
for tier in range(1,6):
s = 0.6*(1-ps[tier]) + 0.4*(TIER_COST[tier]/TIER_COST[5])
if s < best_s: best_s = s; best = tier
return best, ps
eval_router("CARROT_v1_mu0.6", carrot_v1)
# 6. Safety-first CARROT (mu=0.7, threshold=0.5)
def safe_carrot_050(x, t):
return route_safe_carrot(x, tier_clfs, t["tt"], mu=0.7, success_threshold=0.5)
eval_router("safe_CARROT_t0.50", safe_carrot_050)
# 7. Safety-first CARROT (mu=0.7, threshold=0.6)
def safe_carrot_060(x, t):
return route_safe_carrot(x, tier_clfs, t["tt"], mu=0.7, success_threshold=0.6)
eval_router("safe_CARROT_t0.60", safe_carrot_060)
# 8. Safety-first CARROT (mu=0.7, threshold=0.65)
def safe_carrot_065(x, t):
return route_safe_carrot(x, tier_clfs, t["tt"], mu=0.7, success_threshold=0.65)
eval_router("safe_CARROT_t0.65", safe_carrot_065)
# 9. Oracle
eval_router("oracle", lambda x, t: (t["opt"], {}))
# Print comparison
print(f"\n{'Router':<25} {'Success':>10} {'AvgCost':>10} {'CostRed':>10} {'Unsafe':>10} {'F-DONE':>10}")
print("-"*75)
frontier_cost = results["always_frontier"]["avg_cost"]
for name, r in sorted(results.items(), key=lambda x: -x[1]["success"]):
cr = (1 - r["avg_cost"]/frontier_cost)*100
print(f"{name:<25} {r['success']:>10.3f} {r['avg_cost']:>10.4f} {cr:>9.1f}% {r['unsafe_rate']:>10.3f} {r['false_done']:>10.3f}")
# ─── Train Improved Direct Classifier ───────────────────────────────
print("\n\n[5] Training improved direct classifier (0-indexed)...")
y_train_direct = y_opt[idx_train] - 1
y_test_direct = y_opt[idx_test] - 1
# Use sample weights: penalize underprediction more
from sklearn.utils.class_weight import compute_sample_weight
# Custom weight: underkill is 3x worse than overkill
sample_weights = []
for i in idx_train:
t = traces[i]
opt = t["opt"]
# Weight by inverse frequency + safety penalty
sample_weights.append(1.0)
sample_weights = np.array(sample_weights)
direct_clf = xgb.XGBClassifier(
n_estimators=300, max_depth=6, learning_rate=0.05,
subsample=0.8, colsample_bytree=0.8,
objective="multi:softmax", num_class=5,
eval_metric="mlogloss", random_state=42, verbosity=0,
)
direct_clf.fit(X_train, y_train_direct, sample_weight=sample_weights)
y_pred_direct = direct_clf.predict(X_test) + 1 # back to 1-indexed
acc = accuracy_score(y_opt[idx_test], y_pred_direct)
print(f" Direct classifier accuracy: {acc:.3f}")
# Evaluate direct classifier with safety floors
def direct_safe(x, t):
pred = int(direct_clf.predict(x)[0]) + 1
floor = TASK_FLOOR.get(t["tt"], 2)
return max(pred, floor), {}
eval_router("direct_safe_xgb", direct_safe)
# ─── Feature Importance ─────────────────────────────────────────────
print("\n\n[6] Feature importance (from direct classifier)...")
imp = direct_clf.feature_importances_
for feat, score in sorted(zip(FEAT_KEYS, imp), key=lambda x: -x[1])[:10]:
print(f" {feat:<25}: {score:.4f}")
# ─── Save Models ────────────────────────────────────────────────────
print("\n\n[7] Saving models...")
os.makedirs("/app/router_models", exist_ok=True)
for tier, clf in tier_clfs.items():
clf.save_model(f"/app/router_models/tier_{tier}_success.json")
direct_clf.save_model("/app/router_models/direct_optimal_tier.json")
with open("/app/router_models/feat_keys.json", "w") as f:
json.dump(FEAT_KEYS, f)
with open("/app/router_models/tier_config.json", "w") as f:
json.dump({"tier_cost": TIER_COST, "tier_str": TIER_STR, "task_floor": TASK_FLOOR}, f, indent=2)
# Final print
print(f"\n\n{'='*80}")
print("FINAL COMPARISON (ALL ROUTERS)")
print(f"{'='*80}")
print(f"\n{'Router':<25} {'Success':>10} {'AvgCost':>10} {'CostRed':>10} {'Unsafe':>10} {'F-DONE':>10}")
print("-"*75)
frontier_cost = results["always_frontier"]["avg_cost"]
for name, r in sorted(results.items(), key=lambda x: (-x[1]["success"], x[1]["avg_cost"])):
cr = (1 - r["avg_cost"]/frontier_cost)*100
print(f"{name:<25} {r['success']:>10.3f} {r['avg_cost']:>10.4f} {cr:>9.1f}% {r['unsafe_rate']:>10.3f} {r['false_done']:>10.3f}")
print(f"\n\nDONE! Models saved to /app/router_models/")
# ─── RouteLLM-Style Binary Router ────────────────────────────────────
print("\n\n[8] Training RouteLLM-style binary classifiers...")
print(" (For each tier pair, train: should we route to cheaper or more expensive tier?)")
# For each tier boundary, train a binary classifier
# tier_boundary[t] = P(should use tier >= t | query)
# Route to the first tier where the boundary classifier says "yes, this is enough"
boundary_clfs = {}
for boundary in range(2, 6):
# Label: 1 if optimal_tier < boundary (cheaper tier is sufficient)
# 0 if optimal_tier >= boundary (need this tier or higher)
y_boundary = np.array([1 if traces[i]["opt"] < boundary else 0 for i in range(len(traces))])
y_tr = y_boundary[idx_train]
y_te = y_boundary[idx_test]
neg = (y_tr == 0).sum()
pos = (y_tr == 1).sum()
spw = neg / max(pos, 1)
clf = xgb.XGBClassifier(
n_estimators=150, max_depth=5, learning_rate=0.1,
subsample=0.8, colsample_bytree=0.8,
scale_pos_weight=min(spw, 3.0),
objective="binary:logistic", eval_metric="logloss",
random_state=42, verbosity=0,
)
clf.fit(X_train, y_tr)
y_pred = clf.predict(X_test)
acc = accuracy_score(y_te, y_pred)
f1 = f1_score(y_te, y_pred, zero_division=0)
boundary_clfs[boundary] = clf
rate = (y_tr == 0).mean() # fraction that needs this tier
print(f" Boundary {boundary}: acc={acc:.3f}, f1={f1:.3f}, needs_tier={rate:.3f}")
def route_cascade_binary(x, t):
"""RouteLLM-style cascade: check each boundary, route to first that passes."""
if x.ndim == 1:
x = x.reshape(1, -1)
floor = TASK_FLOOR.get(t["tt"], 2)
# Start at floor, check if we need higher
current_tier = floor
for boundary in range(floor + 1, 6):
# boundary_clfs[boundary] predicts P(optimal < boundary)
# If P(optimal < boundary) > threshold, we can stay below boundary
# i.e., if P(need tier >= boundary) > threshold, escalate
p_need_higher = boundary_clfs[boundary].predict_proba(x)[0, 0] # P(optimal >= boundary)
if p_need_higher > 0.4: # confidence threshold
current_tier = boundary
else:
break
return current_tier, {}
eval_router("cascade_binary_t0.4", route_cascade_binary)
def route_cascade_binary_t050(x, t):
if x.ndim == 1: x = x.reshape(1, -1)
floor = TASK_FLOOR.get(t["tt"], 2)
current_tier = floor
for boundary in range(floor + 1, 6):
p_need = boundary_clfs[boundary].predict_proba(x)[0, 0]
if p_need > 0.5:
current_tier = boundary
else:
break
return current_tier, {}
eval_router("cascade_binary_t0.5", route_cascade_binary_t050)
def route_cascade_binary_t030(x, t):
if x.ndim == 1: x = x.reshape(1, -1)
floor = TASK_FLOOR.get(t["tt"], 2)
current_tier = floor
for boundary in range(floor + 1, 6):
p_need = boundary_clfs[boundary].predict_proba(x)[0, 0]
if p_need > 0.3:
current_tier = boundary
else:
break
return current_tier, {}
eval_router("cascade_binary_t0.3", route_cascade_binary_t030)
# Save boundary classifiers
for boundary, clf in boundary_clfs.items():
clf.save_model(f"/app/router_models/boundary_{boundary}.json")
print(f" Saved boundary_{boundary}.json")
# ─── Final Final Comparison ───────────────────────────────────────────
print(f"\n\n{'='*80}")
print("FINAL COMPARISON v2 (WITH BINARY CASCADE ROUTER)")
print(f"{'='*80}")
print(f"\n{'Router':<25} {'Success':>10} {'AvgCost':>10} {'CostRed':>10} {'Unsafe':>10} {'F-DONE':>10}")
print("-"*75)
frontier_cost = results["always_frontier"]["avg_cost"]
for name, r in sorted(results.items(), key=lambda x: (-x[1]["success"], x[1]["avg_cost"])):
cr = (1 - r["avg_cost"]/frontier_cost)*100
# Only show key results
if name in ("oracle","always_frontier","heuristic_diff+1","safe_CARROT_t0.60",
"cascade_binary_t0.4","cascade_binary_t0.5","cascade_binary_t0.3",
"always_cheap"):
print(f"{name:<25} {r['success']:>10.3f} {r['avg_cost']:>10.4f} {cr:>9.1f}% {r['unsafe_rate']:>10.3f} {r['false_done']:>10.3f}")
# Find best Pareto
print("\n\nPARETO FRONTIER:")
pareto = []
for name, r in results.items():
if name in ("always_cheap",):
continue # skip dominated
dominated = False
for name2, r2 in results.items():
if name == name2: continue
if r2["success"] >= r["success"] and r2["avg_cost"] <= r["avg_cost"]:
if r2["success"] > r["success"] or r2["avg_cost"] < r["avg_cost"]:
dominated = True; break
if not dominated:
pareto.append((name, r))
cr = (1 - r["avg_cost"]/frontier_cost)*100
print(f" {name:<25} success={r['success']:.3f} cost={r['avg_cost']:.4f} costRed={cr:.1f}%")
# Save all results
with open("/app/router_models/eval_results.json", "w") as f:
json.dump(results, f, indent=2, default=str)
print(f"\n Saved eval_results.json")
print(f"\nDONE!")