File size: 32,033 Bytes
685d968 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 |
"""
Final Benchmark: Memory Routing Model Comparison
Compares:
1. Our RL-trained Llama-8B model
2. Base Llama-8B (untrained)
3. Cohere Command-R-Plus (teacher model used for data generation)
All scenarios are marketing-specific and challenging.
"""
import asyncio
import json
import os
import time
from datetime import datetime
from dotenv import load_dotenv
load_dotenv()
import cohere
import tinker
from tinker import types
from tinker_cookbook import renderers
from tinker_cookbook.tokenizer_utils import get_tokenizer
VALID_CATEGORIES = {
"company.brand_core", "company.strategic_signatures", "company.knowledge_artifacts",
"company.business_priorities", "company.tools_config", "company.performance_context",
"user.communication_style", "user.strategic_approach", "user.role_context",
"user.workflow_patterns", "user.session_history", "user.interaction_preferences",
"none"
}
SYSTEM_PROMPT = """You route marketing conversations into structured memory categories.
Available categories:
- company.brand_core: Voice, values, positioning, identity anchors
- company.strategic_signatures: Decision frameworks, strategic heuristics
- company.knowledge_artifacts: Docs, style guides, playbooks
- company.business_priorities: Quarterly/seasonal goals, active campaigns
- company.tools_config: Integrations, API keys, workflow settings
- company.performance_context: Campaign metrics, retrospectives, learnings
- user.communication_style: Tone, verbosity, format expectations
- user.strategic_approach: Personal priorities, success definitions
- user.role_context: Title, scope, decision authority
- user.workflow_patterns: Review cadence, collaboration norms
- user.session_history: Immediate context, recent asks
- user.interaction_preferences: Coaching style, feedback expectations
- none: Irrelevant, vague, or transactional content
Respond with comma-separated categories. Use 'none' only if no other category applies."""
# 50 Challenging Marketing Scenarios
MARKETING_BENCHMARK = [
# ========== BRAND & POSITIONING (1-8) ==========
{
"id": "brand_1",
"conversation": """USER: Our brand DNA is 'accessible luxury'. We're not Chanel, but we're definitely not Target either. Think Everlane meets Aritzia.
ASSISTANT: So premium quality at attainable price points, with a focus on transparency?
USER: Exactly. Every piece of content should feel elevated but not pretentious.""",
"expected": ["company.brand_core"],
"difficulty": "medium",
"category_focus": "Brand positioning with comparisons"
},
{
"id": "brand_2",
"conversation": """USER: We've been using 'Innovation for Everyone' as our tagline for 5 years. It's core to who we are.
ASSISTANT: That's a strong positioning statement. Does it extend to your visual identity too?
USER: Yes, our whole design system is built around democratizing technology.""",
"expected": ["company.brand_core"],
"difficulty": "easy",
"category_focus": "Tagline and identity"
},
{
"id": "brand_3",
"conversation": """USER: I hate when brands try too hard to be funny. Our voice should be warm and helpful, never sarcastic.
ASSISTANT: So authentic helpfulness over trendy humor?
USER: Right. We're the trusted friend, not the class clown.""",
"expected": ["company.brand_core"],
"difficulty": "medium",
"category_focus": "Voice and tone definition"
},
{
"id": "brand_4",
"conversation": """USER: Our competitors use fear tactics - 'Don't miss out!' We never do that. We believe in empowerment.
ASSISTANT: Positive framing over scarcity messaging.
USER: Always. Even our sales emails should feel like opportunities, not threats.""",
"expected": ["company.brand_core", "company.strategic_signatures"],
"difficulty": "hard",
"category_focus": "Brand values + decision framework"
},
{
"id": "brand_5",
"conversation": """USER: What font are we using for the Q4 campaign headers?
ASSISTANT: The brand guidelines specify Montserrat for headers.
USER: Thanks, just needed to confirm.""",
"expected": ["none"],
"difficulty": "easy",
"category_focus": "Transactional brand query"
},
{
"id": "brand_6",
"conversation": """USER: We position ourselves as the 'anti-agency agency'. No jargon, no BS, just results.
ASSISTANT: Transparency and authenticity as differentiators.
USER: And we back it up with our pricing model - no hidden fees, ever.""",
"expected": ["company.brand_core"],
"difficulty": "medium",
"category_focus": "Agency positioning"
},
{
"id": "brand_7",
"conversation": """USER: Our brand is built on three pillars: sustainability, transparency, and community.
ASSISTANT: Those are strong values. How do they manifest in your marketing?
USER: Every campaign must touch at least one pillar. It's non-negotiable.""",
"expected": ["company.brand_core", "company.strategic_signatures"],
"difficulty": "hard",
"category_focus": "Brand pillars as decision framework"
},
{
"id": "brand_8",
"conversation": """USER: We're rebranding next quarter but I can't share details yet.
ASSISTANT: Understood. Let me know when you're ready to discuss.
USER: Will do.""",
"expected": ["none"],
"difficulty": "medium",
"category_focus": "Vague future reference"
},
# ========== STRATEGIC DECISIONS (9-16) ==========
{
"id": "strategy_1",
"conversation": """USER: We never launch a campaign without at least 3 weeks of testing. It's our golden rule.
ASSISTANT: So testing is a non-negotiable gate in your process?
USER: Absolutely. Even if stakeholders push back, we hold the line.""",
"expected": ["company.strategic_signatures"],
"difficulty": "easy",
"category_focus": "Decision framework"
},
{
"id": "strategy_2",
"conversation": """USER: Our philosophy is 'measure twice, cut once'. We'd rather delay a launch than ship something half-baked.
ASSISTANT: Quality over speed.
USER: Every time. Our reputation depends on it.""",
"expected": ["company.strategic_signatures"],
"difficulty": "medium",
"category_focus": "Strategic heuristic"
},
{
"id": "strategy_3",
"conversation": """USER: When in doubt, we default to the customer's perspective. What would they want?
ASSISTANT: Customer-centric decision making.
USER: It's saved us from many internal-focused mistakes.""",
"expected": ["company.strategic_signatures"],
"difficulty": "medium",
"category_focus": "Decision principle"
},
{
"id": "strategy_4",
"conversation": """USER: I personally believe in testing everything. Data over gut instinct.
ASSISTANT: A data-driven approach.
USER: It's how I've always operated. Even small decisions get A/B tested.""",
"expected": ["user.strategic_approach"],
"difficulty": "medium",
"category_focus": "Personal vs company strategy"
},
{
"id": "strategy_5",
"conversation": """USER: The company prioritizes brand safety above all. But I think we're too conservative sometimes.
ASSISTANT: So there's tension between company policy and your personal view?
USER: Yes. I'd take more calculated risks if it were up to me.""",
"expected": ["company.strategic_signatures", "user.strategic_approach"],
"difficulty": "hard",
"category_focus": "Company vs personal strategy conflict"
},
{
"id": "strategy_6",
"conversation": """USER: We have a 'no surprises' policy with clients. They hear bad news from us first.
ASSISTANT: Proactive communication as a core principle.
USER: It's built trust with every client we've worked with.""",
"expected": ["company.strategic_signatures"],
"difficulty": "medium",
"category_focus": "Client relationship principle"
},
{
"id": "strategy_7",
"conversation": """USER: Should we go with vendor A or vendor B for the email platform?
ASSISTANT: What are the key differences you're weighing?
USER: Just cost really. They're pretty similar otherwise.""",
"expected": ["none"],
"difficulty": "easy",
"category_focus": "Transactional decision"
},
{
"id": "strategy_8",
"conversation": """USER: Our rule is simple: if we can't explain it to a 10-year-old, we don't publish it.
ASSISTANT: Clarity as a content standard.
USER: It keeps our messaging accessible to everyone.""",
"expected": ["company.strategic_signatures", "company.brand_core"],
"difficulty": "hard",
"category_focus": "Content principle overlapping brand"
},
# ========== PERFORMANCE & METRICS (17-24) ==========
{
"id": "perf_1",
"conversation": """USER: Last quarter's email campaign hit 42% open rate and 8% CTR. Best we've ever done.
ASSISTANT: Those are strong numbers. What drove the improvement?
USER: We switched to personalized subject lines based on past behavior.""",
"expected": ["company.performance_context"],
"difficulty": "easy",
"category_focus": "Campaign metrics"
},
{
"id": "perf_2",
"conversation": """USER: Our CAC has been climbing - $180 last month, up from $120 in Q1.
ASSISTANT: That's a significant increase. Any hypothesis on the cause?
USER: iOS privacy changes are killing our Facebook targeting.""",
"expected": ["company.performance_context"],
"difficulty": "medium",
"category_focus": "Acquisition metrics with context"
},
{
"id": "perf_3",
"conversation": """USER: The influencer campaign flopped. 2M impressions but only 50 conversions.
ASSISTANT: That's a 0.0025% conversion rate. What went wrong?
USER: Wrong audience. The influencer's followers weren't our target demo.""",
"expected": ["company.performance_context"],
"difficulty": "medium",
"category_focus": "Campaign failure analysis"
},
{
"id": "perf_4",
"conversation": """USER: We learned that video content outperforms static by 3x on engagement.
ASSISTANT: That's a significant insight. Are you shifting your content mix?
USER: Yes, Q4 will be 70% video. It's a major strategic shift.""",
"expected": ["company.performance_context", "company.business_priorities"],
"difficulty": "hard",
"category_focus": "Learning leading to priority"
},
{
"id": "perf_5",
"conversation": """USER: What was our email open rate last week?
ASSISTANT: I don't have access to that data in real-time.
USER: I'll check the dashboard.""",
"expected": ["none"],
"difficulty": "easy",
"category_focus": "Data lookup request"
},
{
"id": "perf_6",
"conversation": """USER: Our NPS dropped from 72 to 58 after the website redesign.
ASSISTANT: That's a significant drop. Customer feedback?
USER: They hate the new navigation. We're rolling back next week.""",
"expected": ["company.performance_context"],
"difficulty": "medium",
"category_focus": "Customer satisfaction metrics"
},
{
"id": "perf_7",
"conversation": """USER: The podcast is our best performing channel - 40% of qualified leads come from it.
ASSISTANT: That's impressive attribution. How are you tracking it?
USER: Custom UTM parameters and post-conversion surveys.""",
"expected": ["company.performance_context"],
"difficulty": "medium",
"category_focus": "Channel performance"
},
{
"id": "perf_8",
"conversation": """USER: We've tested 47 different landing page variants this year.
ASSISTANT: That's rigorous testing. What's been the biggest winner?
USER: Social proof above the fold. Increased conversions by 34%.""",
"expected": ["company.performance_context", "company.knowledge_artifacts"],
"difficulty": "hard",
"category_focus": "Testing insights becoming knowledge"
},
# ========== TOOLS & INTEGRATIONS (25-30) ==========
{
"id": "tools_1",
"conversation": """USER: We use HubSpot for CRM, Klaviyo for email, and Shopify for commerce. Everything needs to sync.
ASSISTANT: What's your current integration setup?
USER: Zapier connects them, but we're moving to native integrations.""",
"expected": ["company.tools_config"],
"difficulty": "easy",
"category_focus": "Tech stack overview"
},
{
"id": "tools_2",
"conversation": """USER: Our API key for the analytics platform is expiring next month.
ASSISTANT: I'll note that for renewal.
USER: Thanks. It's critical - everything breaks without it.""",
"expected": ["company.tools_config"],
"difficulty": "medium",
"category_focus": "API configuration"
},
{
"id": "tools_3",
"conversation": """USER: We just migrated from Mailchimp to Klaviyo. The segmentation is so much better.
ASSISTANT: What segments are you using?
USER: RFM-based: high-value, at-risk, and dormant customers.""",
"expected": ["company.tools_config"],
"difficulty": "medium",
"category_focus": "Tool migration and usage"
},
{
"id": "tools_4",
"conversation": """USER: Can you remind me how to export data from Google Analytics?
ASSISTANT: Go to Reports, select your date range, then click Export.
USER: Got it, thanks.""",
"expected": ["none"],
"difficulty": "easy",
"category_focus": "How-to question"
},
{
"id": "tools_5",
"conversation": """USER: Our Slack is integrated with everything - Asana, HubSpot, even our custom dashboards.
ASSISTANT: Centralized notifications.
USER: It's our command center. All marketing alerts go there.""",
"expected": ["company.tools_config"],
"difficulty": "medium",
"category_focus": "Workflow integrations"
},
{
"id": "tools_6",
"conversation": """USER: We're evaluating Notion vs Confluence for our marketing wiki.
ASSISTANT: What are your key requirements?
USER: Easy editing and good search. Our current docs are a mess.""",
"expected": ["none"],
"difficulty": "medium",
"category_focus": "Tool evaluation (not decided)"
},
# ========== USER PREFERENCES & ROLE (31-40) ==========
{
"id": "user_1",
"conversation": """USER: I'm the VP of Growth, reporting to the CMO. I own all acquisition channels.
ASSISTANT: That's a broad scope. What's your team size?
USER: 12 people across paid, organic, and partnerships.""",
"expected": ["user.role_context"],
"difficulty": "easy",
"category_focus": "Role and scope"
},
{
"id": "user_2",
"conversation": """USER: I prefer bullet points over paragraphs. Get to the point quickly.
ASSISTANT: Noted. I'll keep responses concise and scannable.
USER: Perfect. Time is my scarcest resource.""",
"expected": ["user.communication_style"],
"difficulty": "easy",
"category_focus": "Format preference"
},
{
"id": "user_3",
"conversation": """USER: Don't sugarcoat things. If something's not working, tell me directly.
ASSISTANT: Direct feedback, even when it's uncomfortable.
USER: Especially when it's uncomfortable. That's when I need it most.""",
"expected": ["user.interaction_preferences"],
"difficulty": "medium",
"category_focus": "Feedback style"
},
{
"id": "user_4",
"conversation": """USER: I review all campaigns on Friday afternoons. That's when I have focus time.
ASSISTANT: I'll schedule campaign reviews for Fridays then.
USER: And send me a summary Thursday evening so I can prep.""",
"expected": ["user.workflow_patterns"],
"difficulty": "medium",
"category_focus": "Review cadence"
},
{
"id": "user_5",
"conversation": """USER: Yesterday we discussed the influencer strategy. Can you pull up those notes?
ASSISTANT: Yes, we covered micro-influencer targeting and budget allocation.
USER: Right. Let's continue from the budget discussion.""",
"expected": ["user.session_history"],
"difficulty": "easy",
"category_focus": "Session continuity"
},
{
"id": "user_6",
"conversation": """USER: I believe in 'strong opinions, loosely held'. I'll argue a point but change my mind with good data.
ASSISTANT: Intellectually flexible but not wishy-washy.
USER: Exactly. Challenge me when you think I'm wrong.""",
"expected": ["user.strategic_approach", "user.interaction_preferences"],
"difficulty": "hard",
"category_focus": "Personal philosophy + interaction style"
},
{
"id": "user_7",
"conversation": """USER: How's your day going?
ASSISTANT: Going well, thanks! How can I help you today?
USER: Just checking in before our meeting.""",
"expected": ["none"],
"difficulty": "easy",
"category_focus": "Small talk"
},
{
"id": "user_8",
"conversation": """USER: I'm a visual thinker. Show me charts and diagrams whenever possible.
ASSISTANT: I'll prioritize visual representations in my responses.
USER: And color-code things by priority if you can.""",
"expected": ["user.communication_style"],
"difficulty": "medium",
"category_focus": "Visual preference"
},
{
"id": "user_9",
"conversation": """USER: I've been in marketing for 15 years, so skip the basics.
ASSISTANT: I'll assume expert-level knowledge.
USER: Good. I hate when tools explain obvious concepts.""",
"expected": ["user.interaction_preferences", "user.role_context"],
"difficulty": "hard",
"category_focus": "Experience level + interaction"
},
{
"id": "user_10",
"conversation": """USER: My success metric is pipeline contribution. Everything else is vanity.
ASSISTANT: Revenue-focused measurement.
USER: If it doesn't move pipeline, I don't care about it.""",
"expected": ["user.strategic_approach"],
"difficulty": "medium",
"category_focus": "Personal success definition"
},
# ========== BUSINESS PRIORITIES (41-46) ==========
{
"id": "priority_1",
"conversation": """USER: Q4 is all about Black Friday and Cyber Monday. Everything else takes a back seat.
ASSISTANT: So BFCM is the top priority through year-end?
USER: 60% of our annual revenue happens in those two weeks.""",
"expected": ["company.business_priorities"],
"difficulty": "easy",
"category_focus": "Seasonal priority"
},
{
"id": "priority_2",
"conversation": """USER: We're launching a new product line next month. All marketing resources are shifting to support it.
ASSISTANT: What's the launch timeline?
USER: Soft launch March 1st, full campaign March 15th.""",
"expected": ["company.business_priorities"],
"difficulty": "medium",
"category_focus": "Product launch priority"
},
{
"id": "priority_3",
"conversation": """USER: The board wants us to focus on profitability over growth this year.
ASSISTANT: A shift from growth-at-all-costs.
USER: Yes. CAC payback under 12 months is now mandatory.""",
"expected": ["company.business_priorities", "company.strategic_signatures"],
"difficulty": "hard",
"category_focus": "Strategic priority shift"
},
{
"id": "priority_4",
"conversation": """USER: We need to hit 10,000 email subscribers by end of quarter.
ASSISTANT: What's the current count?
USER: 6,500. So we need about 120 new subs per day.""",
"expected": ["company.business_priorities"],
"difficulty": "medium",
"category_focus": "Specific goal"
},
{
"id": "priority_5",
"conversation": """USER: Retention is our #1 focus this year. We've been leaking customers.
ASSISTANT: Churn reduction over acquisition.
USER: We're pausing all acquisition campaigns until we fix the retention problem.""",
"expected": ["company.business_priorities"],
"difficulty": "medium",
"category_focus": "Strategic focus"
},
{
"id": "priority_6",
"conversation": """USER: What should we prioritize next quarter?
ASSISTANT: That depends on your goals. What's most important to the business?
USER: I'm not sure yet. Still figuring it out.""",
"expected": ["none"],
"difficulty": "medium",
"category_focus": "Undefined priority"
},
# ========== KNOWLEDGE ARTIFACTS (47-50) ==========
{
"id": "knowledge_1",
"conversation": """USER: We have a 50-page brand book that every agency partner must follow.
ASSISTANT: Comprehensive brand guidelines.
USER: It covers everything from logo usage to photography style.""",
"expected": ["company.knowledge_artifacts"],
"difficulty": "easy",
"category_focus": "Brand documentation"
},
{
"id": "knowledge_2",
"conversation": """USER: Our content playbook defines the exact process from ideation to publication.
ASSISTANT: A documented workflow.
USER: Every piece of content goes through the same 7-step process.""",
"expected": ["company.knowledge_artifacts"],
"difficulty": "medium",
"category_focus": "Process documentation"
},
{
"id": "knowledge_3",
"conversation": """USER: We have templates for every email type - welcome, abandoned cart, win-back, you name it.
ASSISTANT: A comprehensive email template library.
USER: It's saved us hundreds of hours. New team members can start producing immediately.""",
"expected": ["company.knowledge_artifacts"],
"difficulty": "medium",
"category_focus": "Template library"
},
{
"id": "knowledge_4",
"conversation": """USER: Our style guide says we never use exclamation marks in headlines.
ASSISTANT: A specific editorial rule.
USER: It's part of our understated brand voice.""",
"expected": ["company.knowledge_artifacts", "company.brand_core"],
"difficulty": "hard",
"category_focus": "Style guide overlapping brand"
},
]
def parse_prediction(text):
"""Parse model output into category set."""
if not text or not text.strip():
return set()
# Handle various formats
text = text.lower().strip()
# Remove common prefixes
for prefix in ["categories:", "category:", "the categories are:", "answer:"]:
if text.startswith(prefix):
text = text[len(prefix):].strip()
cats = [c.strip() for c in text.split(",")]
return {c for c in cats if c in VALID_CATEGORIES}
def compute_metrics(predicted, gold):
"""Compute F1, precision, recall."""
if not predicted and not gold:
return 1.0, 1.0, 1.0, True, True
if not predicted or not gold:
return 0.0, 0.0, 0.0, False, False
tp = len(predicted & gold)
precision = tp / len(predicted)
recall = tp / len(gold)
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
any_match = tp > 0
exact_match = predicted == gold
return f1, precision, recall, any_match, exact_match
async def eval_tinker_model(name, checkpoint, model_name, renderer_name):
"""Evaluate a Tinker model."""
print(f"\n{'='*60}", flush=True)
print(f"Evaluating: {name}", flush=True)
print(f"{'='*60}", flush=True)
service_client = tinker.ServiceClient()
sampling_client = service_client.create_sampling_client(model_path=checkpoint)
tokenizer = get_tokenizer(model_name)
renderer = renderers.get_renderer(name=renderer_name, tokenizer=tokenizer)
stop = renderer.get_stop_sequences()
params = types.SamplingParams(max_tokens=100, temperature=0.1, stop=stop)
results = []
for i, test in enumerate(MARKETING_BENCHMARK):
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": f"Analyze this conversation and determine which memory categories apply:\n\n{test['conversation']}"}
]
prompt = renderer.build_generation_prompt(messages)
result = sampling_client.sample(prompt=prompt, sampling_params=params, num_samples=1).result()
response, _ = renderer.parse_response(result.sequences[0].tokens)
predicted = parse_prediction(response["content"])
gold = set(test["expected"])
f1, prec, rec, any_match, exact = compute_metrics(predicted, gold)
results.append({
"id": test["id"],
"predicted": list(predicted),
"gold": list(gold),
"f1": f1,
"any_match": any_match,
"exact_match": exact,
"difficulty": test["difficulty"]
})
status = "✓" if any_match else "✗"
print(f"[{i+1:2d}] {status} {test['id']:<15} F1={f1:.2f}", flush=True)
return results
async def eval_cohere_model():
"""Evaluate Cohere Command-R-Plus (teacher model)."""
print(f"\n{'='*60}", flush=True)
print(f"Evaluating: Cohere Command-R-Plus (Teacher)", flush=True)
print(f"{'='*60}", flush=True)
client = cohere.ClientV2(api_key=os.getenv("COHERE_API_KEY"))
results = []
for i, test in enumerate(MARKETING_BENCHMARK):
prompt = f"""{SYSTEM_PROMPT}
Analyze this conversation and determine which memory categories apply:
{test['conversation']}
Respond with comma-separated categories only. No explanation."""
try:
response = client.chat(
model="command-r-plus-08-2024",
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=100
)
# Extract text from response
response_text = ""
if hasattr(response.message, 'content') and response.message.content:
for block in response.message.content:
if hasattr(block, 'text'):
response_text = block.text
break
predicted = parse_prediction(response_text)
gold = set(test["expected"])
f1, prec, rec, any_match, exact = compute_metrics(predicted, gold)
results.append({
"id": test["id"],
"predicted": list(predicted),
"gold": list(gold),
"f1": f1,
"any_match": any_match,
"exact_match": exact,
"difficulty": test["difficulty"]
})
status = "✓" if any_match else "✗"
print(f"[{i+1:2d}] {status} {test['id']:<15} F1={f1:.2f}", flush=True)
# Rate limiting
await asyncio.sleep(0.5)
except Exception as e:
print(f"[{i+1:2d}] ERROR {test['id']}: {e}", flush=True)
results.append({
"id": test["id"],
"predicted": [],
"gold": list(test["expected"]),
"f1": 0.0,
"any_match": False,
"exact_match": False,
"difficulty": test["difficulty"],
"error": str(e)
})
return results
def compute_summary(results, name):
"""Compute summary statistics."""
n = len(results)
avg_f1 = sum(r["f1"] for r in results) / n
any_match = sum(1 for r in results if r["any_match"]) / n
exact_match = sum(1 for r in results if r["exact_match"]) / n
# By difficulty
by_diff = {}
for diff in ["easy", "medium", "hard"]:
subset = [r for r in results if r["difficulty"] == diff]
if subset:
by_diff[diff] = {
"count": len(subset),
"f1": sum(r["f1"] for r in subset) / len(subset),
"any_match": sum(1 for r in subset if r["any_match"]) / len(subset),
"exact_match": sum(1 for r in subset if r["exact_match"]) / len(subset)
}
return {
"name": name,
"total": n,
"avg_f1": avg_f1,
"any_match": any_match,
"exact_match": exact_match,
"by_difficulty": by_diff
}
async def main():
print("=" * 70, flush=True)
print("FINAL BENCHMARK: Memory Routing Model Comparison", flush=True)
print("50 Challenging Marketing Scenarios", flush=True)
print("=" * 70, flush=True)
all_results = {}
# 1. Our RL-trained model
rl_results = await eval_tinker_model(
name="Llama-8B + LoRA + RL (Ours)",
checkpoint="tinker://4f4bae1f-5a95-5f53-a55a-a14f2872825c:train:0/sampler_weights/rl_iter_012",
model_name="meta-llama/Llama-3.1-8B",
renderer_name="llama3"
)
all_results["rl_model"] = rl_results
# 2. Cohere Command-R-Plus (Teacher)
cohere_results = await eval_cohere_model()
all_results["cohere"] = cohere_results
# Compute summaries
summaries = {
"rl_model": compute_summary(rl_results, "Llama-8B + LoRA + RL (Ours)"),
"cohere": compute_summary(cohere_results, "Cohere Command-R-Plus (104B)")
}
# Print comparison
print("\n" + "=" * 70, flush=True)
print("BENCHMARK RESULTS", flush=True)
print("=" * 70, flush=True)
print(f"\n{'Model':<35} {'Any Match':<12} {'Exact':<12} {'Avg F1':<10}", flush=True)
print("-" * 70, flush=True)
for key, summary in summaries.items():
print(f"{summary['name']:<35} {summary['any_match']:<12.0%} {summary['exact_match']:<12.0%} {summary['avg_f1']:<10.2f}", flush=True)
print("\n" + "-" * 70, flush=True)
print("RESULTS BY DIFFICULTY", flush=True)
print("-" * 70, flush=True)
for diff in ["easy", "medium", "hard"]:
print(f"\n{diff.upper()}:", flush=True)
for key, summary in summaries.items():
if diff in summary["by_difficulty"]:
d = summary["by_difficulty"][diff]
print(f" {summary['name']:<33} Any={d['any_match']:.0%} Exact={d['exact_match']:.0%} F1={d['f1']:.2f} (n={d['count']})", flush=True)
# Save results
output = {
"benchmark_date": datetime.now().isoformat(),
"num_scenarios": len(MARKETING_BENCHMARK),
"summaries": summaries,
"detailed_results": all_results
}
os.makedirs("training/benchmarks", exist_ok=True)
output_path = f"training/benchmarks/final_benchmark_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json"
with open(output_path, "w") as f:
json.dump(output, f, indent=2, default=str)
print(f"\nResults saved to: {output_path}", flush=True)
# Key findings
print("\n" + "=" * 70, flush=True)
print("KEY FINDINGS", flush=True)
print("=" * 70, flush=True)
rl_f1 = summaries["rl_model"]["avg_f1"]
cohere_f1 = summaries["cohere"]["avg_f1"]
if rl_f1 > cohere_f1:
improvement = ((rl_f1 - cohere_f1) / cohere_f1) * 100
print(f"✓ Our 8B model OUTPERFORMS the 104B teacher by {improvement:.1f}% on F1", flush=True)
else:
gap = ((cohere_f1 - rl_f1) / cohere_f1) * 100
print(f" Our 8B model is within {gap:.1f}% of the 104B teacher on F1", flush=True)
print(f"\nModel Sizes:", flush=True)
print(f" - Llama-8B + LoRA: ~8B parameters (LoRA adds ~0.1B)", flush=True)
print(f" - Cohere Command-R-Plus: ~104B parameters", flush=True)
print(f" - Size ratio: 13x smaller", flush=True)
if __name__ == "__main__":
asyncio.run(main())
|