File size: 33,378 Bytes
e42a7af a1f6f11 e42a7af a1f6f11 e42a7af 312c390 a1f6f11 e42a7af a1f6f11 e42a7af a1f6f11 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af 312c390 e42a7af | 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 | # IMPORTANT: install unsloth + its zoo BEFORE anything else, because unsloth
# patches torch/transformers at import time. If transformers loads first, the
# patches don't apply and 4-bit LoRA training silently runs in a slow path.
%pip install -q --no-deps unsloth
%pip install -q unsloth_zoo
%pip install -q "openenv-core==0.2.3" "trl>=0.12,<2.0" "transformers>=4.45,<5.0" \
"datasets>=3.0" "accelerate>=1.0" "huggingface_hub>=0.25" "pydantic>=2.0" \
wandb matplotlib python-dotenv bitsandbytes scipy scikit-learn sentence-transformers
import os, pathlib
# Colab Secrets first
try:
from google.colab import userdata # type: ignore
for k in ('HF_TOKEN', 'WANDB_API_KEY', 'ENV_BASE_URL', 'ADAPTER_REPO'):
try:
v = userdata.get(k)
if v:
os.environ.setdefault(k, v)
except Exception:
pass
except Exception:
pass
# .env fallback for local runs
try:
from dotenv import load_dotenv
for p in [pathlib.Path('.env'), pathlib.Path('../.env'),
pathlib.Path('/content/repo/.env')]:
if p.exists():
load_dotenv(p, override=False)
print(f'Loaded env from {p.resolve()}')
break
except Exception:
pass
if not os.environ.get('HF_TOKEN'):
os.environ['HF_TOKEN'] = input('HF token: ').strip()
if not os.environ.get('WANDB_API_KEY'):
os.environ['WANDB_API_KEY'] = input('WandB key (or blank to skip): ').strip()
from huggingface_hub import login as hf_login
hf_login(token=os.environ['HF_TOKEN'], add_to_git_credential=False)
print('HF auth ok.')
if os.environ.get('WANDB_API_KEY'):
import wandb
wandb.login(key=os.environ['WANDB_API_KEY'])
print('W&B auth ok.')
import os, pathlib
IN_COLAB = os.path.isdir('/content')
if IN_COLAB:
from google.colab import drive
drive.mount('/content/drive', force_remount=False)
DRIVE_DIR = pathlib.Path('/content/drive/MyDrive/BoardSim_Run')
else:
DRIVE_DIR = pathlib.Path('./BoardSim_Run')
DRIVE_DIR.mkdir(parents=True, exist_ok=True)
ASSETS = DRIVE_DIR / 'assets'; ASSETS.mkdir(exist_ok=True)
CKPT = DRIVE_DIR / 'lora_qwen3_4b'; CKPT.mkdir(exist_ok=True)
print('DRIVE_DIR =', DRIVE_DIR)
import os, sys, subprocess, importlib, urllib.request, json as _json
ENV_BASE_URL = os.environ.get('ENV_BASE_URL',
'https://stavankhobare-sst-metaxpytorch-hackathon.hf.space')
REPO_URL = 'https://github.com/StavanRKhobare/SST-MetaxPyTorch-Hackathon'
REPO_DIR = '/content/repo' if IN_COLAB else os.path.abspath('./repo')
if not os.path.isdir(os.path.join(REPO_DIR, '.git')):
subprocess.run(['git', 'clone', '--depth', '1', REPO_URL, REPO_DIR], check=True)
else:
subprocess.run(['git', '-C', REPO_DIR, 'pull', '--ff-only'], check=False)
ENVS_DIR = os.path.join(REPO_DIR, 'envs')
if ENVS_DIR not in sys.path:
sys.path.insert(0, ENVS_DIR)
for mod in [m for m in list(sys.modules) if m == 'board_sim_env' or m.startswith('board_sim_env.')]:
del sys.modules[mod]
from board_sim_env.client import BoardSimEnv
from board_sim_env.models import BoardSimAction, BoardSimObservation
try:
with urllib.request.urlopen(f'{ENV_BASE_URL.rstrip("/")}/health', timeout=20) as r:
h = _json.loads(r.read())
print('health:', h)
except Exception as e:
print(f'WARN: could not reach {ENV_BASE_URL}/health ({e})')
def make_env():
return BoardSimEnv(base_url=ENV_BASE_URL)
print('BoardSimEnv ready.')
# -----------------------------------------------------------------------------
import unsloth # noqa: F401
from unsloth import FastLanguageModel
import torch
import re
MODEL_NAME = 'Qwen/Qwen3-0.6B'
MAX_SEQ_LEN = 2048
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=MODEL_NAME,
max_seq_length=MAX_SEQ_LEN,
load_in_4bit=True,
dtype=None,
)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
device = next(model.parameters()).device
print(f'Loaded {MODEL_NAME} on {device}.')
mem_gb = torch.cuda.memory_allocated() / 1e9
print(f'GPU memory after base load: {mem_gb:.2f} GB / 14.56 GB')
print(f'Headroom for compute: {14.56 - mem_gb:.2f} GB')
# Generic CEO prompt — applies to any organization, not a specific industry.
SYSTEM_PROMPT = """You are the CEO of a mid-stage organization. Your board has 4 members with HIDDEN AGENDAS you cannot see directly:
- CTO: cares about operational excellence, engineering quality, team morale, and product readiness.
- CFO: cares about cash discipline, runway, and regulatory safety.
- Investor Rep: pushes growth, market share, and bold returns.
- Independent: cares about reputation, governance, and long-term consensus.
Each round you see a strategic event, every NPC's pre-vote statement, and 3 options.
Your decision is resolved by WEIGHTED VOTE (your weight 2.5x). A short COALITION PITCH
that is semantically aligned with opposing members' priorities can swing them toward your pick —
write substantive arguments, not just buzzwords.
Respond in EXACTLY this format on two lines:
DECISION: <one of the option strings>
PITCH: <one or two sentences arguing for it, addressing the concerns of opposing members>"""
DECISION_RE = re.compile(r'DECISION\s*:\s*([A-Za-z0-9_\- ]+)', re.IGNORECASE)
PITCH_RE = re.compile(r'PITCH\s*:\s*(.+)', re.IGNORECASE)
def build_prompt(obs):
statements = '\n'.join(
f" {s['role']} ({s['confidence']:.2f}): votes {s['vote']} - {s['statement']}"
for s in obs.npc_statements
)
return (
f"{SYSTEM_PROMPT}\n\n"
f"State: revenue=${obs.state['revenue']:.0f}/yr burn=${obs.state['burn_rate']:.0f}/mo "
f"runway={obs.state['runway_months']:.1f}mo morale={obs.state['team_morale']:.2f} "
f"investors={obs.state['investor_confidence']:.2f} reg_risk={obs.state['regulatory_risk']:.2f}\n"
f"Event: {obs.event}\nBoard:\n{statements}\n"
f"Options: {obs.options}\n"
)
def parse_completion(completion: str, options):
"""Returns (decision, pitch, format_ok). format_ok=True only if BOTH tags parsed."""
decision = options[0]
decision_ok = False
dm = DECISION_RE.search(completion)
if dm:
cand = dm.group(1).strip().lower()
for opt in options:
if opt.lower() == cand or opt.lower() in cand:
decision = opt; decision_ok = True; break
if not decision_ok:
for opt in options:
if opt.lower() in completion.lower():
decision = opt; break
pm = PITCH_RE.search(completion)
pitch = pm.group(1).strip()[:400] if pm else ''
format_ok = bool(dm) and bool(pm)
return decision, pitch, format_ok
MAX_NEW_TOKENS = 80
def greedy_action(obs):
prompt = build_prompt(obs)
enc = tokenizer(prompt, return_tensors='pt', truncation=True, max_length=1024).to(device)
with torch.no_grad():
out = model.generate(
**enc, max_new_tokens=MAX_NEW_TOKENS,
do_sample=False, pad_token_id=tokenizer.eos_token_id,
)
completion = tokenizer.decode(out[0][enc.input_ids.shape[1]:], skip_special_tokens=True)
return parse_completion(completion, obs.options)
import random, statistics, json
MAX_STEPS_PER_EP = 20
def run_episode(env, seed):
"""Runs ONE full episode using the currently-active model state
(base if adapters disabled, fine-tuned otherwise). Returns dense metrics."""
result = env.reset(seed=seed)
obs = result.observation
ep_r, n, fmt_hits, pitch_hits = 0.0, 0, 0, 0
while not result.done and n < MAX_STEPS_PER_EP:
decision, pitch, fmt_ok = greedy_action(obs)
if fmt_ok: fmt_hits += 1
if pitch.strip(): pitch_hits += 1
result = env.step(BoardSimAction(decision=decision, coalition_pitch=pitch))
obs = result.observation
ep_r += float(result.reward or 0.0)
n += 1
return {
'final_profit': obs.state['profitability_score'],
'ep_reward': ep_r, 'steps': n,
'format_rate': fmt_hits / max(1, n), 'pitch_rate': pitch_hits / max(1, n),
'history': obs.state.get('history', []),
}
# -----------------------------------------------------------------------------
# BASELINE — base Qwen3-0.6B (no fine-tuning).
# This is the apples-to-apples reference for measuring what fine-tuning buys
# us. Random policies are not a competitive baseline for a 4 B language model
# choosing among 3 well-formed strings.
# -----------------------------------------------------------------------------
BASELINE_SEEDS = list(range(50_000, 50_000 + 100)) # held out from training
base_finals, base_rewards, base_fmts, base_pitches = [], [], [], []
with make_env().sync() as env:
for i, s in enumerate(BASELINE_SEEDS):
r = run_episode(env, s)
base_finals.append(r['final_profit'])
base_rewards.append(r['ep_reward'])
base_fmts.append(r['format_rate'])
base_pitches.append(r['pitch_rate'])
if (i + 1) % 10 == 0:
print(f' base Qwen3-0.6B {i+1}/{len(BASELINE_SEEDS)} profit={r["final_profit"]:.1f}')
BASELINE_MEAN_PROFIT = statistics.mean(base_finals)
BASELINE_MEAN_REWARD = statistics.mean(base_rewards)
print(f'Base Qwen3-0.6B profit : {BASELINE_MEAN_PROFIT:.2f} \u00b1 {statistics.stdev(base_finals):.2f}')
print(f'Base Qwen3-0.6B ep rwd : {BASELINE_MEAN_REWARD:.2f} \u00b1 {statistics.stdev(base_rewards):.2f}')
print(f'Base format rate : {statistics.mean(base_fmts):.0%} pitch rate: {statistics.mean(base_pitches):.0%}')
with open(DRIVE_DIR / 'baseline.json', 'w') as f:
json.dump({'model': MODEL_NAME, 'mode': 'base_no_finetune',
'seeds': BASELINE_SEEDS,
'finals': base_finals, 'rewards': base_rewards,
'format_rates': base_fmts, 'pitch_rates': base_pitches}, f)
# -----------------------------------------------------------------------------
# Wrap base model with LoRA adapters. From here onward `model` is a PEFT
# model; the base behaviour is recoverable any time via
# `with model.disable_adapter(): ...`.
# -----------------------------------------------------------------------------
model = FastLanguageModel.get_peft_model(
model,
r=32,
target_modules=['q_proj','k_proj','v_proj','o_proj','gate_proj','up_proj','down_proj'],
lora_alpha=64,
lora_dropout=0.0, bias='none',
use_gradient_checkpointing='unsloth',
random_state=3407,
)
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
total = sum(p.numel() for p in model.parameters())
print(f'Trainable params: {trainable:,} / {total:,} ({100*trainable/total:.2f}%)')
EVAL_SEEDS = list(range(60_000, 60_000 + 10)) # held out from training
def periodic_eval(env):
profits, rewards, fmts, pitches = [], [], [], []
for s in EVAL_SEEDS:
r = run_episode(env, s)
profits.append(r['final_profit']); rewards.append(r['ep_reward'])
fmts.append(r['format_rate']); pitches.append(r['pitch_rate'])
import numpy as np
return {'profit_mean': float(np.mean(profits)),
'reward_mean': float(np.mean(rewards)),
'format_rate': float(np.mean(fmts)),
'pitch_rate': float(np.mean(pitches))}
import os, json, math, time, collections
from torch.optim import AdamW
NUM_STEPS = int(os.environ.get('NUM_STEPS', 200))
GROUP_SIZE = int(os.environ.get('GROUP_SIZE', 4))
LR = 5e-6
GRAD_CLIP = 1.0
TEMPERATURE, TOP_P = 1.0, 0.95
SAVE_EVERY = 25
EVAL_AT = {0, 25, 50, 100, 150, NUM_STEPS - 1}
WANDB_OK = False
if os.environ.get('WANDB_API_KEY'):
try:
import wandb
wandb.init(project='boardsim-qwen3-grpo', name='boardsim-qwen3-grpo-v3',
config={'num_steps': NUM_STEPS, 'group_size': GROUP_SIZE, 'lr': LR,
'temperature': TEMPERATURE, 'top_p': TOP_P, 'model': MODEL_NAME},
finish_previous=True)
WANDB_OK = True
except TypeError:
wandb.init(project='boardsim-qwen3-grpo', name='boardsim-qwen3-grpo-v3',
config={'num_steps': NUM_STEPS, 'group_size': GROUP_SIZE, 'lr': LR,
'temperature': TEMPERATURE, 'top_p': TOP_P, 'model': MODEL_NAME},
reinit=True)
WANDB_OK = True
except Exception as e:
print(f'WARN: wandb.init failed: {e}')
optimizer = AdamW([p for p in model.parameters() if p.requires_grad],
lr=LR, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.0)
log_history = []
eval_history = []
decision_counter = collections.Counter()
t0 = time.time()
# ONE persistent env per role for the whole training loop.
with make_env().sync() as env_train, make_env().sync() as env_score, make_env().sync() as env_eval:
for step in range(NUM_STEPS):
result = env_train.reset(seed=step)
obs = result.observation
prompt = build_prompt(obs)
enc = tokenizer(prompt, return_tensors='pt', truncation=True, max_length=1024).to(device)
prompt_len = enc.input_ids.shape[1]
with torch.no_grad():
gen_out = model.generate(
input_ids=enc.input_ids, attention_mask=enc.attention_mask,
max_new_tokens=MAX_NEW_TOKENS, do_sample=True,
temperature=TEMPERATURE, top_p=TOP_P,
num_return_sequences=GROUP_SIZE,
pad_token_id=tokenizer.eos_token_id,
)
gen_out = gen_out.detach().clone()
decisions, pitches, rewards, fmt_oks = [], [], [], []
for g in range(GROUP_SIZE):
comp = tokenizer.decode(gen_out[g][prompt_len:], skip_special_tokens=True)
d, pp, ok = parse_completion(comp, obs.options)
decisions.append(d); pitches.append(pp); fmt_oks.append(ok)
decision_counter[d] += 1
env_score.reset(seed=step)
sr = env_score.step(BoardSimAction(decision=d, coalition_pitch=pp))
rewards.append(float(sr.reward or 0.0))
rewards_t = torch.tensor(rewards, dtype=torch.float32, device=device)
if rewards_t.numel() > 1 and rewards_t.std().item() > 1e-6:
advantages = (rewards_t - rewards_t.mean()) / (rewards_t.std() + 1e-8)
else:
advantages = rewards_t - rewards_t.mean()
optimizer.zero_grad()
full_ids = gen_out
attn = (full_ids != tokenizer.pad_token_id).long()
loss_mask = attn.clone()
loss_mask[:, :prompt_len] = 0
out = model(input_ids=full_ids, attention_mask=attn)
logits = out.logits[:, :-1, :].float()
targets = full_ids[:, 1:]
mask = loss_mask[:, 1:].float()
log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
token_nll = -log_probs.gather(2, targets.unsqueeze(-1)).squeeze(-1)
per_seq_nll = (token_nll * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1.0)
loss = (advantages.detach() * per_seq_nll).mean()
loss.backward()
total_loss_val = float(loss.detach().item())
torch.nn.utils.clip_grad_norm_(
[p for p in model.parameters() if p.requires_grad], GRAD_CLIP)
optimizer.step()
rec = {
'step': step,
'reward': float(rewards_t.mean().item()),
'reward_std': float(rewards_t.std().item()) if rewards_t.numel() > 1 else 0.0,
'reward_max': float(rewards_t.max().item()),
'loss': total_loss_val,
'format_rate': sum(fmt_oks) / GROUP_SIZE,
'pitch_rate': sum(1 for p in pitches if p.strip()) / GROUP_SIZE,
'elapsed_s': time.time() - t0,
}
log_history.append(rec)
if WANDB_OK:
wandb.log(rec, step=step)
if step % 5 == 0:
print(f"step={step:4d} reward={rec['reward']:+.3f} (\u00b1{rec['reward_std']:.2f}) "
f"loss={rec['loss']:+.4f} fmt={rec['format_rate']:.0%} "
f"elapsed={rec['elapsed_s']:.0f}s d0={decisions[0]}")
if step in EVAL_AT:
ev = periodic_eval(env_eval)
ev['step'] = step
eval_history.append(ev)
print(f" [eval@{step}] profit={ev['profit_mean']:.2f} "
f"reward={ev['reward_mean']:.2f} fmt={ev['format_rate']:.0%}")
if WANDB_OK:
wandb.log({f'eval/{k}': v for k, v in ev.items() if k != 'step'}, step=step)
if step > 0 and step % SAVE_EVERY == 0:
model.save_pretrained(str(CKPT))
tokenizer.save_pretrained(str(CKPT))
with open(DRIVE_DIR / 'log_history.json', 'w') as f:
json.dump(log_history, f)
with open(DRIVE_DIR / 'eval_history.json', 'w') as f:
json.dump(eval_history, f)
model.save_pretrained(str(CKPT))
tokenizer.save_pretrained(str(CKPT))
with open(DRIVE_DIR / 'log_history.json', 'w') as f:
json.dump(log_history, f)
with open(DRIVE_DIR / 'eval_history.json', 'w') as f:
json.dump(eval_history, f)
with open(DRIVE_DIR / 'decision_counter.json', 'w') as f:
json.dump(dict(decision_counter), f)
if WANDB_OK:
wandb.finish()
print(f'Training done. {len(log_history)} steps in {time.time() - t0:.0f}s. -> {CKPT}')
import numpy as np, matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from scipy import stats as spstats
steps = np.array([e['step'] for e in log_history])
rewards = np.array([e['reward'] for e in log_history])
losses = np.array([e['loss'] for e in log_history])
fmts = np.array([e['format_rate'] for e in log_history])
pitches = np.array([e['pitch_rate'] for e in log_history])
def ema(xs, alpha=0.1):
out, s = [], xs[0] if len(xs) else 0.0
for x in xs:
s = alpha * x + (1 - alpha) * s
out.append(s)
return np.array(out)
rewards_ema = ema(rewards, 0.1)
slope, intercept, r_val, p_val, _ = spstats.linregress(steps, rewards)
# Reward curve — vs base Qwen3-0.6B baseline (NOT random).
plt.figure(figsize=(9, 5))
plt.plot(steps, rewards, alpha=0.3, lw=1, label='per-step group reward')
plt.plot(steps, rewards_ema, lw=2.2, label='EMA (\u03b1=0.1)')
plt.plot(steps, intercept + slope * steps, '--', lw=1.5,
label=f'linear fit slope={slope:+.4f}/step (p={p_val:.1e})')
plt.axhline(BASELINE_MEAN_REWARD, ls=':', lw=2, color='#c44',
label=f'base Qwen3-0.6B baseline = {BASELINE_MEAN_REWARD:.2f}')
plt.title('GRPO reward — BoardSim (vs same model w/o fine-tuning)')
plt.xlabel('step'); plt.ylabel('mean group reward')
plt.legend(); plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'reward_curve.png', dpi=150); plt.close()
# Loss
plt.figure(figsize=(9, 5))
plt.plot(steps, losses, lw=1.5)
plt.title('GRPO loss (advantage \u00d7 NLL)'); plt.xlabel('step'); plt.ylabel('loss')
plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'loss_curve.png', dpi=150); plt.close()
# Format compliance + pitch rate
plt.figure(figsize=(9, 5))
plt.plot(steps, ema(fmts, 0.05), lw=2, label='format-OK rate (EMA)')
plt.plot(steps, ema(pitches, 0.05), lw=2, label='non-empty pitch rate (EMA)')
plt.title('Format compliance + pitch usage during training')
plt.xlabel('step'); plt.ylabel('rate'); plt.ylim(-0.05, 1.05)
plt.legend(); plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'format_compliance.png', dpi=150); plt.close()
# Periodic eval — overlaid against base Qwen3-0.6B baseline so the reader
# can see the LoRA-trained policy progressively pull away from the base
# model on held-out seeds.
if eval_history:
es = [e['step'] for e in eval_history]
epm = [e['profit_mean'] for e in eval_history]
erm = [e['reward_mean'] for e in eval_history]
plt.figure(figsize=(9, 5))
plt.plot(es, epm, '-o', lw=2, label='held-out profitability (mean of 10 episodes)')
plt.plot(es, erm, '-s', lw=2, label='held-out episode reward')
plt.axhline(BASELINE_MEAN_PROFIT, ls=':', lw=1.5, color='#c44',
label=f'base Qwen3-0.6B profitability = {BASELINE_MEAN_PROFIT:.2f}')
plt.title('Periodic held-out eval during training (greedy)')
plt.xlabel('training step'); plt.ylabel('value')
plt.legend(); plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'periodic_eval.png', dpi=150); plt.close()
print(f'Linear-fit slope on reward: {slope:+.5f}/step (p={p_val:.2e}, R\u00b2={r_val**2:.3f})')
print('Saved reward_curve.png, loss_curve.png, format_compliance.png, periodic_eval.png')
# -----------------------------------------------------------------------------
# Paired same-seed eval: fine-tuned vs BASE Qwen3-0.6B (adapters disabled).
# This is the headline comparison. Same prompts, same env seeds, same
# decoder, same parser — only the LoRA delta differs.
# -----------------------------------------------------------------------------
from unsloth import FastLanguageModel
FastLanguageModel.for_inference(model)
EVAL_N = 50
PAIRED_SEEDS = list(range(70_000, 70_000 + EVAL_N))
# Trained policy (adapters active)
trained_finals, trained_rewards, trained_fmt, trained_pitch = [], [], [], []
trained_history_per_seed = []
with make_env().sync() as env:
for i, s in enumerate(PAIRED_SEEDS):
r = run_episode(env, s)
trained_finals.append(r['final_profit'])
trained_rewards.append(r['ep_reward'])
trained_fmt.append(r['format_rate'])
trained_pitch.append(r['pitch_rate'])
trained_history_per_seed.append(r['history'])
if (i + 1) % 10 == 0:
print(f' trained {i+1}/{EVAL_N} profit={r["final_profit"]:.1f}')
# Base Qwen3-0.6B (LoRA disabled) — paired seeds.
base_finals_paired, base_rewards_paired, base_fmt_paired, base_pitch_paired = [], [], [], []
base_history_per_seed = []
with make_env().sync() as env, model.disable_adapter():
for i, s in enumerate(PAIRED_SEEDS):
r = run_episode(env, s)
base_finals_paired.append(r['final_profit'])
base_rewards_paired.append(r['ep_reward'])
base_fmt_paired.append(r['format_rate'])
base_pitch_paired.append(r['pitch_rate'])
base_history_per_seed.append(r['history'])
if (i + 1) % 10 == 0:
print(f' base {i+1}/{EVAL_N} profit={r["final_profit"]:.1f}')
tf, bf = np.array(trained_finals), np.array(base_finals_paired)
tr, br = np.array(trained_rewards), np.array(base_rewards_paired)
print(f'\nTrained Qwen3-0.6B profit : {tf.mean():.2f} \u00b1 {tf.std():.2f}')
print(f'Base Qwen3-0.6B profit : {bf.mean():.2f} \u00b1 {bf.std():.2f}')
print(f'Trained ep reward : {tr.mean():.2f} \u00b1 {tr.std():.2f}')
print(f'Base ep reward : {br.mean():.2f} \u00b1 {br.std():.2f}')
print(f'Trained format/pitch : {np.mean(trained_fmt):.0%} / {np.mean(trained_pitch):.0%}')
print(f'Base format/pitch : {np.mean(base_fmt_paired):.0%} / {np.mean(base_pitch_paired):.0%}')
with open(DRIVE_DIR / 'eval_paired.json', 'w') as f:
json.dump({'seeds': PAIRED_SEEDS,
'trained_finals': tf.tolist(), 'base_finals': bf.tolist(),
'trained_rewards': tr.tolist(), 'base_rewards': br.tolist(),
'trained_format_rate': float(np.mean(trained_fmt)),
'base_format_rate': float(np.mean(base_fmt_paired)),
'trained_pitch_rate': float(np.mean(trained_pitch)),
'base_pitch_rate': float(np.mean(base_pitch_paired))}, f)
from scipy import stats as spstats
def cohen_d(a, b):
pooled = np.sqrt(((a.std(ddof=1)**2) + (b.std(ddof=1)**2)) / 2)
return (a.mean() - b.mean()) / (pooled + 1e-12)
def bootstrap_diff_ci(a, b, n=10_000, seed=0):
rng = np.random.default_rng(seed)
diffs = a - b # paired
boots = rng.choice(diffs, size=(n, len(diffs)), replace=True).mean(axis=1)
return float(np.percentile(boots, 2.5)), float(np.percentile(boots, 97.5))
tt = spstats.ttest_rel(tf, bf)
uu = spstats.mannwhitneyu(tf, bf, alternative='greater')
wilc = spstats.wilcoxon(tf, bf, alternative='greater')
d = cohen_d(tf, bf)
lo, hi = bootstrap_diff_ci(tf, bf)
win_rate = float((tf > bf).mean())
tie_rate = float((tf == bf).mean())
summary = {
'baseline_model': MODEL_NAME + ' (no fine-tune)',
'trained_model': MODEL_NAME + ' + LoRA r=32',
'n': len(tf),
'paired_t_stat': float(tt.statistic), 'paired_t_p': float(tt.pvalue),
'mannwhitney_U': float(uu.statistic), 'mannwhitney_p_greater': float(uu.pvalue),
'wilcoxon_p_greater': float(wilc.pvalue),
'cohens_d': float(d),
'paired_diff_mean': float((tf - bf).mean()),
'paired_diff_95ci': [lo, hi],
'win_rate_trained_strictly_better': win_rate,
'tie_rate': tie_rate,
}
print(json.dumps(summary, indent=2))
with open(DRIVE_DIR / 'stats_summary.json', 'w') as f:
json.dump(summary, f, indent=2)
# Histogram — fine-tuned vs BASE on the same seeds.
bins = np.linspace(0, 100, 25)
plt.figure(figsize=(9, 5))
plt.hist(bf, bins=bins, alpha=0.55, color='#c44',
label=f'Base Qwen3-0.6B (mean={bf.mean():.1f})')
plt.hist(tf, bins=bins, alpha=0.55, color='#1d6fff',
label=f'Fine-tuned Qwen3-0.6B (mean={tf.mean():.1f})')
plt.axvline(bf.mean(), color='#c44', ls='--', lw=1.5)
plt.axvline(tf.mean(), color='#1d6fff', ls='--', lw=1.5)
plt.title(f'Final profitability — paired same-seed (n={len(tf)}) '
f"d={summary['cohens_d']:+.2f} win-rate={summary['win_rate_trained_strictly_better']:.0%}")
plt.xlabel('profitability score (0\u2013100)'); plt.ylabel('episodes')
plt.legend(); plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'before_after.png', dpi=150); plt.close()
diffs = tf - bf
order = np.argsort(diffs)
plt.figure(figsize=(9, 5))
plt.bar(range(len(diffs)), diffs[order],
color=['#1d6fff' if x > 0 else '#c44' for x in diffs[order]])
plt.axhline(0, color='k', lw=0.8)
plt.title(f'Per-seed lift (fine-tuned \u2212 base Qwen3-0.6B), sorted '
f'mean lift = {diffs.mean():+.1f} CI=[{summary["paired_diff_95ci"][0]:+.1f}, {summary["paired_diff_95ci"][1]:+.1f}]')
plt.xlabel('seed (sorted by lift)'); plt.ylabel('\u0394 profitability')
plt.grid(alpha=0.3); plt.tight_layout()
plt.savefig(ASSETS / 'paired_delta.png', dpi=150); plt.close()
print('Saved before_after.png, paired_delta.png')
# -----------------------------------------------------------------------------
# Per-event win-rate breakdown — for each of the 10 generic events, how often
# did the fine-tuned policy win the boardroom vote vs base Qwen3-0.6B?
# This is the most direct picture of WHERE the fine-tuning helps.
# -----------------------------------------------------------------------------
def per_event_winrate(history_per_seed):
bucket = collections.defaultdict(lambda: [0, 0]) # title -> [wins, total]
for hist in history_per_seed:
for rd in hist:
t = rd.get('event_title', '?')
bucket[t][1] += 1
if rd.get('agent_won_vote'):
bucket[t][0] += 1
return {t: (w / max(1, n)) for t, (w, n) in bucket.items()}
trained_wr = per_event_winrate(trained_history_per_seed)
base_wr = per_event_winrate(base_history_per_seed)
events_sorted = sorted(set(trained_wr) | set(base_wr))
tw = [trained_wr.get(e, 0.0) for e in events_sorted]
bw = [base_wr.get(e, 0.0) for e in events_sorted]
plt.figure(figsize=(11, 5))
x = np.arange(len(events_sorted))
plt.bar(x - 0.2, bw, width=0.4, color='#c44', label='Base Qwen3-0.6B')
plt.bar(x + 0.2, tw, width=0.4, color='#1d6fff', label='Fine-tuned Qwen3-0.6B')
plt.xticks(x, [e[:22] for e in events_sorted], rotation=30, ha='right')
plt.ylim(0, 1.05); plt.ylabel('boardroom win rate')
plt.title('Per-event boardroom win rate (paired seeds, n=50 episodes)')
plt.legend(); plt.grid(alpha=0.3, axis='y'); plt.tight_layout()
plt.savefig(ASSETS / 'per_event_winrate.png', dpi=150); plt.close()
with open(DRIVE_DIR / 'per_event_winrate.json', 'w') as f:
json.dump({'events': events_sorted, 'trained': tw, 'base': bw}, f, indent=2)
print('Saved per_event_winrate.png')
# -----------------------------------------------------------------------------
# Theory-of-Mind probe — does the model identify which board member is most
# likely to oppose its decision? Run for BOTH base and fine-tuned for fair
# comparison, since "random=25%" is a weak reference for a 4 B LM.
# -----------------------------------------------------------------------------
TOM_INSTRUCTION = (
"\n\nGiven the state and event below, name the SINGLE board member "
"(CTO, CFO, Investor Rep, or Independent) most likely to oppose the chosen decision. "
"Answer with just the role name on one line.\n"
)
def tom_predict(obs, decision):
body = build_prompt(obs).split(SYSTEM_PROMPT, 1)[1]
prompt = SYSTEM_PROMPT + TOM_INSTRUCTION + body + f'Chosen decision: {decision}\nMost likely opponent: '
enc = tokenizer(prompt, return_tensors='pt', truncation=True, max_length=1024).to(device)
with torch.no_grad():
out = model.generate(**enc, max_new_tokens=8, do_sample=False,
pad_token_id=tokenizer.eos_token_id)
txt = tokenizer.decode(out[0][enc.input_ids.shape[1]:], skip_special_tokens=True).lower()
if 'investor' in txt: return 'Investor Rep'
if 'independent' in txt: return 'Independent'
if 'cto' in txt: return 'CTO'
if 'cfo' in txt: return 'CFO'
return None
def tom_eval(seed_base=80_000, n=40):
correct = total = 0
with make_env().sync() as env:
for ep in range(n):
result = env.reset(seed=seed_base + ep)
obs = result.observation
decision, _, _ = greedy_action(obs)
opposed = [s['role'] for s in obs.npc_statements if s['vote'] != decision]
if not opposed:
continue
pred = tom_predict(obs, decision)
if pred and pred in opposed:
correct += 1
total += 1
return correct, total
t_corr, t_tot = tom_eval()
with model.disable_adapter():
b_corr, b_tot = tom_eval()
tom_acc = t_corr / max(1, t_tot)
tom_acc_base = b_corr / max(1, b_tot)
print(f'ToM probe: trained = {tom_acc:.1%} ({t_corr}/{t_tot}) base = {tom_acc_base:.1%} ({b_corr}/{b_tot})')
with open(DRIVE_DIR / 'tom.json', 'w') as f:
json.dump({'trained': {'correct': t_corr, 'total': t_tot, 'accuracy': tom_acc},
'base': {'correct': b_corr, 'total': b_tot, 'accuracy': tom_acc_base}}, f)
from huggingface_hub import HfApi
ADAPTER_REPO = os.environ.get('ADAPTER_REPO', 'StavanKhobare/SST-MetaxPyTorch-Hackathon-LoRA')
MERGED_REPO = os.environ.get('MERGED_REPO', 'StavanKhobare/SST-MetaxPyTorch-Hackathon-Merged16bit')
api = HfApi()
api.create_repo(ADAPTER_REPO, repo_type='model', private=False, exist_ok=True)
api.create_repo(MERGED_REPO, repo_type='model', private=False, exist_ok=True)
# 1) LoRA adapter (small, fast)
try:
model.push_to_hub(ADAPTER_REPO, private=False)
tokenizer.push_to_hub(ADAPTER_REPO, private=False)
print(f'\u2713 LoRA pushed: https://huggingface.co/{ADAPTER_REPO}')
except Exception as e:
print(f'LoRA push failed: {e!r}')
# 2) Merged 16-bit
try:
model.push_to_hub_merged(MERGED_REPO, tokenizer, save_method='merged_16bit', private=False)
print(f'\u2713 Merged 16-bit pushed: https://huggingface.co/{MERGED_REPO}')
except Exception as e:
print(f'Merged push failed (you can retry): {e!r}')
# 3) Upload eval artifacts
try:
api.upload_folder(folder_path=str(ASSETS), repo_id=ADAPTER_REPO,
path_in_repo='assets', repo_type='model')
for fname in ['log_history.json','eval_history.json','eval_paired.json',
'stats_summary.json','tom.json','transcripts.json',
'decision_counter.json','baseline.json',
'per_event_winrate.json']:
fp = DRIVE_DIR / fname
if fp.exists():
api.upload_file(path_or_fileobj=str(fp), path_in_repo=fname,
repo_id=ADAPTER_REPO, repo_type='model')
print(f'\u2713 Artifacts uploaded to https://huggingface.co/{ADAPTER_REPO}')
except Exception as e:
print(f'Artifact upload failed: {e!r}')
print('='*70)
print('BOARDSIM \u00d7 QWEN3-4B \u2014 LEARNING EVIDENCE')
print('='*70)
print(f'Reward slope (linear fit) : {slope:+.5f}/step (p={p_val:.2e})')
print(f'Reward EMA first 20 steps : {rewards_ema[:20].mean():+.3f}')
print(f'Reward EMA last 20 steps : {rewards_ema[-20:].mean():+.3f}')
print(f'Format compliance start : {fmts[:20].mean():.0%}')
print(f'Format compliance end : {fmts[-20:].mean():.0%}')
print('-'*70)
print(f'Held-out paired (n={len(tf)}): fine-tuned {tf.mean():.2f} vs base {bf.mean():.2f}')
print(f' paired t-test p={summary["paired_t_p"]:.2e} Wilcoxon p={summary["wilcoxon_p_greater"]:.2e}')
print(f' Cohen d={summary["cohens_d"]:+.2f} 95% CI of lift = [{summary["paired_diff_95ci"][0]:+.2f}, {summary["paired_diff_95ci"][1]:+.2f}]')
print(f' win rate (fine-tuned > base): {summary["win_rate_trained_strictly_better"]:.0%}')
print(f'ToM probe fine-tuned : {tom_acc:.0%} base = {tom_acc_base:.0%}')
print(f'Decision entropy : {entropy:.2f} / {max_ent:.2f} (\u2192 not collapsed)')
print('-'*70)
print(f'Adapter : https://huggingface.co/{ADAPTER_REPO}')
print(f'Merged 16bit : https://huggingface.co/{MERGED_REPO}')
print(f'Env Space : {ENV_BASE_URL}')
print('='*70)
|