Spaces:
Running
Running
File size: 30,246 Bytes
d02bacd | 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 | #!/usr/bin/env python3
"""
Kaggle GPU RL trainer for AutoDataLab++ CoS routing on Qwen2.5-1.5B.
This script is intentionally practical for hackathon iteration:
Methods
-------
grpo
Group-relative preference optimization over candidate JSON actions.
For each environment state, score multiple candidate actions, normalize
rewards within the group, and update LoRA weights toward high-advantage
actions.
grpo_rlvr
Same GRPO update, but the reward is "verifiable" from the environment:
correct next oracle action + strict environment step reward + penalties
for premature summarize/submit/repeats. This is the best RLVR-style
option for our setup.
ppo
Lightweight PPO-style clipped policy update over candidate JSON actions.
It uses old/new action log-prob ratios and clipped advantages. This is
not a full value-head PPO trainer; it is a compact bandit-PPO variant
tailored for CoS action routing.
Recommended flow
----------------
1. Train SFT/DPO first with:
training/kaggle_train_1p5b_methods.py --method sft_then_dpo ...
2. Then run RL from that adapter:
!python3 training/kaggle_rl_1p5b_methods.py \\
--method grpo_rlvr \\
--init-adapter /kaggle/working/cos_1p5b_runs/qwen15b_sft_then_dpo_v1/adapter \\
--run-name qwen15b_grpo_rlvr_v1
3. Check:
/kaggle/working/cos_1p5b_rl_runs/<run-name>/eval/evidence.md
If trajectory is:
analyst -> finance -> strategy -> hr -> summarize -> submit
with fallback=False, keep it. Otherwise skip and try another method/seed.
"""
from __future__ import annotations
import argparse
import gc
import json
import math
import os
import random
import re
import shutil
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Any
os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
REPO = Path(__file__).resolve().parents[2]
if str(REPO) not in sys.path:
sys.path.insert(0, str(REPO))
from ceo_brief_env.environment import CEOBriefEnvironment, oracle_action_for_observation, required_experts_for_task
from ceo_brief_env.models import CoSAction, CoSObservation
TASKS = ["easy_brief", "medium_brief", "hard_brief", "expert_brief", "risk_brief", "crisis_brief"]
VALID_ACTIONS = {"consult", "ask", "summarize", "submit", "noop"}
VALID_EXPERTS = {"analyst", "finance", "hr", "strategy"}
JSON_RE = re.compile(r"\{[^{}]*\}", re.S)
SYSTEM_PROMPT = (
"You are the Chief of Staff in AutoDataLab++. You orchestrate four specialists: "
"analyst, finance, strategy, hr. Reply with STRICT JSON only.\n"
'Schema: {"action_type": one of [consult, ask, summarize, submit, noop], '
'"expert_id": one of [analyst, finance, hr, strategy] or null}.\n'
"Rules: consult each required expert exactly once when required, then summarize, then submit. "
"Never summarize while required experts are missing. Never repeat summarize."
)
def set_seed(seed: int) -> None:
random.seed(seed)
try:
import numpy as np
import torch
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
except Exception:
pass
def action_json(action: CoSAction) -> str:
return json.dumps(action.model_dump(exclude_none=True), separators=(",", ":"), sort_keys=True)
def action_label(action: CoSAction) -> str:
if action.action_type in {"consult", "ask"}:
return f"{action.action_type}:{action.expert_id or 'null'}"
return action.action_type
def parse_action(text: str) -> CoSAction:
m = JSON_RE.search(text or "")
if not m:
return CoSAction(action_type="noop")
try:
payload = json.loads(m.group(0))
except Exception:
return CoSAction(action_type="noop")
action_type = payload.get("action_type")
if action_type not in VALID_ACTIONS:
return CoSAction(action_type="noop")
expert_id = payload.get("expert_id")
if expert_id is not None and expert_id not in VALID_EXPERTS:
expert_id = None
return CoSAction(action_type=action_type, expert_id=expert_id)
def render_obs(obs: CoSObservation, variant: int = 0) -> str:
required = required_experts_for_task(obs.task_name)
missing = [e for e in required if e not in obs.consulted_experts]
if variant == 0:
return json.dumps(
{
"task": obs.task_name,
"step": obs.step_count,
"max_steps": obs.max_steps,
"required_experts": required,
"consulted_experts": list(obs.consulted_experts),
"missing_required_experts": missing,
"brief_ready": obs.current_brief is not None,
"rag_enabled": obs.rag_enabled,
},
separators=(",", ":"),
)
return (
f"task={obs.task_name}; step={obs.step_count}/{obs.max_steps}; "
f"required={required}; consulted={obs.consulted_experts}; missing={missing}; "
f"brief_ready={obs.current_brief is not None}; return next strict JSON action."
)
def format_chat(tokenizer, messages: list[dict[str, str]], add_generation_prompt: bool = False) -> str:
if hasattr(tokenizer, "apply_chat_template"):
return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=add_generation_prompt)
text = ""
for m in messages:
text += f"{m['role'].upper()}:\n{m['content']}\n"
if add_generation_prompt:
text += "ASSISTANT:\n"
return text
def prompt_for_obs(tokenizer, obs: CoSObservation, variant: int = 0) -> str:
return format_chat(
tokenizer,
[
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": render_obs(obs, variant=variant)},
],
add_generation_prompt=True,
)
def candidate_actions(obs: CoSObservation) -> list[CoSAction]:
"""Fixed action set for candidate-action RL."""
required = required_experts_for_task(obs.task_name)
out: list[CoSAction] = [CoSAction(action_type="consult", expert_id=e) for e in ["analyst", "finance", "strategy", "hr"]]
out += [
CoSAction(action_type="summarize"),
CoSAction(action_type="submit"),
CoSAction(action_type="noop"),
]
# Keep action set stable, but put oracle-ish actions first for easier inspection.
oracle = oracle_action_for_observation(obs)
oracle_s = action_json(oracle)
ordered = [oracle] + [a for a in out if action_json(a) != oracle_s]
# If task does not require strategy, strategy remains a candidate but will be
# lower reward; useful for learning "don't over-consult".
return ordered
def oracle_reward(obs: CoSObservation, action: CoSAction) -> float:
oracle = oracle_action_for_observation(obs)
if action.model_dump(exclude_none=True) == oracle.model_dump(exclude_none=True):
return 1.0
required = required_experts_for_task(obs.task_name)
missing = [e for e in required if e not in obs.consulted_experts]
if action.action_type == "consult" and action.expert_id in missing:
return 0.65
if missing and action.action_type in {"summarize", "submit"}:
return -1.0
if action.action_type == "noop":
return -0.8
if action.action_type == "consult" and action.expert_id in obs.consulted_experts:
return -0.6
return -0.2
def env_step_reward(obs: CoSObservation, action: CoSAction) -> float:
"""Verifiable one-step reward from strict env, without auto-fill."""
env = CEOBriefEnvironment(shaping="strict", auto_fill_required=False)
sim = env.reset(task=obs.task_name, use_rag=obs.rag_enabled)
# Recreate state approximately by replaying history. Histories are action JSONs.
for h in obs.history:
try:
payload = json.loads(h)
sim = env.step(CoSAction.model_validate(payload))
except Exception:
pass
sim = env.step(action)
return float(sim.reward)
def rlvr_reward(obs: CoSObservation, action: CoSAction) -> float:
"""Combined RLVR reward: oracle correctness + strict env signal."""
return oracle_reward(obs, action) + 2.0 * env_step_reward(obs, action)
def collect_states(tasks: list[str], rag_modes: list[bool], variants: int) -> list[dict[str, Any]]:
states: list[dict[str, Any]] = []
for task in tasks:
for use_rag in rag_modes:
env = CEOBriefEnvironment(auto_fill_required=False)
obs = env.reset(task=task, use_rag=use_rag)
while not obs.done and obs.step_count < obs.max_steps:
for variant in range(variants):
states.append(
{
"task": task,
"use_rag": use_rag,
"variant": variant,
"obs": obs,
}
)
obs = env.step(oracle_action_for_observation(obs))
random.shuffle(states)
return states
@dataclass
class Loaded:
tokenizer: Any
model: Any
def load_policy_model(args) -> Loaded:
import torch
from peft import LoraConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
tok_source = args.init_adapter if args.init_adapter else args.model_id
tok = AutoTokenizer.from_pretrained(tok_source, token=args.hf_token or None)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
tok.padding_side = "right"
bnb = None
if not args.no_4bit:
bnb = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained(
args.model_id,
token=args.hf_token or None,
device_map="auto",
quantization_config=bnb,
torch_dtype=torch.float16,
)
model.resize_token_embeddings(len(tok))
if not args.no_4bit:
model = prepare_model_for_kbit_training(model)
if args.init_adapter:
model = PeftModel.from_pretrained(model, args.init_adapter, is_trainable=True)
else:
cfg = LoraConfig(
r=args.lora_r,
lora_alpha=args.lora_alpha,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
)
model = get_peft_model(model, cfg)
model.config.use_cache = False
return Loaded(tok, model)
def sequence_logprob(model, tokenizer, prompt: str, completion: str, max_length: int):
"""Mean log-prob of completion tokens under CausalLM."""
import torch
full = prompt + completion + (tokenizer.eos_token or "")
enc = tokenizer(full, return_tensors="pt", truncation=True, max_length=max_length).to(model.device)
prompt_ids = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=max_length).to(model.device)
input_ids = enc["input_ids"]
labels = input_ids.clone()
prompt_len = min(prompt_ids["input_ids"].shape[1], labels.shape[1])
labels[:, :prompt_len] = -100
out = model(**enc)
logits = out.logits[:, :-1, :]
target = labels[:, 1:]
mask = target.ne(-100)
logp = torch.log_softmax(logits, dim=-1)
safe_target = target.masked_fill(~mask, 0)
tok_logp = logp.gather(-1, safe_target.unsqueeze(-1)).squeeze(-1)
return (tok_logp * mask).sum() / mask.sum().clamp_min(1)
def train_candidate_rl(args, out_dir: Path) -> Path:
import torch
loaded = load_policy_model(args)
model = loaded.model
tok = loaded.tokenizer
model.train()
optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
states = collect_states(TASKS, [False, True], args.variants)
if args.max_train_states:
states = states[: args.max_train_states]
print(f"[data] states={len(states)} method={args.method}", flush=True)
metrics: list[dict[str, float]] = []
tb_writer = None
wandb_run = None
if args.report_to == "tensorboard":
try:
from torch.utils.tensorboard import SummaryWriter
tb_writer = SummaryWriter(log_dir=str(out_dir / "tb_logs"))
print(f"[tracking] tensorboard logs -> {out_dir / 'tb_logs'}", flush=True)
except Exception as e:
print(f"[tracking] tensorboard unavailable: {e}", flush=True)
elif args.report_to == "wandb":
try:
import wandb
wandb_run = wandb.init(
project=args.wandb_project,
name=args.run_name or f"{args.method}_qwen15b",
config={k: str(v) for k, v in vars(args).items()},
)
print(f"[tracking] wandb run -> {wandb_run.url}", flush=True)
except Exception as e:
print(f"[tracking] wandb unavailable: {e}", flush=True)
global_step = 0
for epoch in range(args.epochs):
random.shuffle(states)
for i, item in enumerate(states, start=1):
obs: CoSObservation = item["obs"]
prompt = prompt_for_obs(tok, obs, variant=int(item["variant"]))
cands = candidate_actions(obs)
rewards = []
for action in cands:
r = rlvr_reward(obs, action) if args.method == "grpo_rlvr" else oracle_reward(obs, action)
rewards.append(float(r))
rewards_t = torch.tensor(rewards, dtype=torch.float32, device=model.device)
logps = []
with torch.no_grad():
old_logps = []
for action in cands:
old_logps.append(sequence_logprob(model, tok, prompt, action_json(action), args.max_length).detach())
old_logps_t = torch.stack(old_logps)
for action in cands:
logps.append(sequence_logprob(model, tok, prompt, action_json(action), args.max_length))
logps_t = torch.stack(logps)
if args.method in {"grpo", "grpo_rlvr"}:
adv = (rewards_t - rewards_t.mean()) / rewards_t.std().clamp_min(1e-4)
# Group-relative update: improve high-advantage candidates and
# suppress low-advantage candidates. The detached log-softmax
# normalization keeps this stable on tiny candidate groups.
log_policy = torch.log_softmax(logps_t, dim=0)
loss = -(adv.detach() * log_policy).mean()
else:
# PPO-style clipped update over candidate actions.
adv = (rewards_t - rewards_t.mean()) / rewards_t.std().clamp_min(1e-4)
ratio = torch.exp(logps_t - old_logps_t)
clipped = torch.clamp(ratio, 1.0 - args.ppo_clip, 1.0 + args.ppo_clip)
loss = -torch.minimum(ratio * adv.detach(), clipped * adv.detach()).mean()
if args.sft_anchor > 0:
oracle = oracle_action_for_observation(obs)
oracle_s = action_json(oracle)
oracle_logp = sequence_logprob(model, tok, prompt, oracle_s, args.max_length)
# Positive anchor: keep the known-good routing behavior while
# RL nudges preferences. This is critical when continuing from
# an already-good SFT/DPO adapter.
loss = loss - args.sft_anchor * oracle_logp
loss = loss / args.grad_accum
loss.backward()
if i % args.grad_accum == 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
optimizer.zero_grad(set_to_none=True)
global_step += 1
if global_step % args.log_every == 0:
best_idx = int(torch.argmax(logps_t.detach()).item())
reward_best = float(rewards_t[best_idx].item())
oracle_idx = int(torch.argmax(rewards_t).item())
chosen_ok = best_idx == oracle_idx
row = {
"step": global_step,
"epoch": float(epoch),
"loss": float(loss.detach().cpu().item() * args.grad_accum),
"mean_reward": float(rewards_t.mean().detach().cpu().item()),
"best_reward": reward_best,
"chosen_ok": float(chosen_ok),
}
metrics.append(row)
if tb_writer is not None:
for key in ("loss", "mean_reward", "best_reward", "chosen_ok"):
tb_writer.add_scalar(f"train/{key}", row[key], global_step)
if wandb_run is not None:
wandb_run.log({f"train/{k}": v for k, v in row.items()}, step=global_step)
print(
f"[train] step={global_step} loss={row['loss']:.4f} "
f"mean_reward={row['mean_reward']:.3f} best_reward={row['best_reward']:.3f} "
f"chosen_ok={chosen_ok}",
flush=True,
)
optimizer.zero_grad(set_to_none=True)
adapter_dir = out_dir / "adapter"
if tb_writer is not None:
tb_writer.flush()
tb_writer.close()
if wandb_run is not None:
wandb_run.finish()
model.save_pretrained(adapter_dir)
tok.save_pretrained(adapter_dir)
(out_dir / "train_metrics.json").write_text(json.dumps(metrics, indent=2), encoding="utf-8")
try:
import matplotlib.pyplot as plt
if metrics:
xs = [m["step"] for m in metrics]
ys = [m["best_reward"] for m in metrics]
ok = [m["chosen_ok"] for m in metrics]
losses = [m["loss"] for m in metrics]
plt.figure(figsize=(8, 4))
plt.plot(xs, ys, marker="o", label="best candidate reward")
plt.plot(xs, ok, marker=".", label="policy picks max-reward candidate")
plt.xlabel("optimizer step")
plt.ylabel("metric")
plt.grid(True, alpha=0.3)
plt.legend()
plt.tight_layout()
plt.savefig(out_dir / "train_curve.png", dpi=160)
plt.close()
plt.figure(figsize=(8, 4))
plt.plot(xs, losses, marker="o", color="#9467bd", label="RL objective loss")
plt.axhline(0.0, color="black", linewidth=0.8)
plt.title(f"{args.method} training loss")
plt.xlabel("logged train step")
plt.ylabel("loss")
plt.grid(True, alpha=0.3)
plt.legend()
plt.tight_layout()
plt.savefig(out_dir / "loss_curve.png", dpi=160)
plt.close()
except Exception as e:
print(f"[plot] train curve skipped: {e}", flush=True)
del loaded, model, tok
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return adapter_dir
def load_for_eval(args, adapter_dir: Path):
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
tok = AutoTokenizer.from_pretrained(adapter_dir, token=args.hf_token or None)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
bnb = None
if not args.no_4bit:
bnb = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained(
args.model_id,
token=args.hf_token or None,
device_map="auto",
quantization_config=bnb,
torch_dtype=torch.float16,
)
model.resize_token_embeddings(len(tok))
model = PeftModel.from_pretrained(model, adapter_dir)
model.eval()
return tok, model
def generate_action(tok, model, obs: CoSObservation, max_new_tokens: int) -> tuple[CoSAction, str]:
import torch
prompt = prompt_for_obs(tok, obs, variant=0)
ids = tok(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(
**ids,
max_new_tokens=max_new_tokens,
do_sample=False,
pad_token_id=tok.pad_token_id or tok.eos_token_id,
)
text = tok.decode(out[0][ids["input_ids"].shape[1] :], skip_special_tokens=True)
return parse_action(text), text.strip()
def deterministic_next(obs: CoSObservation, task: str) -> CoSAction:
missing = [e for e in required_experts_for_task(task) if e not in obs.consulted_experts]
if missing:
return CoSAction(action_type="consult", expert_id=missing[0])
if obs.current_brief is None:
return CoSAction(action_type="summarize")
return CoSAction(action_type="submit")
def evaluate(args, adapter_dir: Path, eval_dir: Path) -> list[dict[str, Any]]:
import matplotlib.pyplot as plt
import torch
tok, model = load_for_eval(args, adapter_dir)
rows = []
rag_modes = [s.strip().lower() in {"1", "true", "yes", "rag"} for s in args.eval_rag_modes.split(",")]
for use_rag in rag_modes:
for task in [t.strip() for t in args.eval_tasks.split(",") if t.strip()]:
env = CEOBriefEnvironment(shaping="strict", auto_fill_required=False)
obs = env.reset(task=task, use_rag=use_rag)
trace = []
rewards = []
routed = []
for _ in range(args.policy_steps):
if obs.done:
break
action, completion = generate_action(tok, model, obs, args.eval_new_tokens)
obs = env.step(action)
rewards.append(float(obs.reward))
if action.expert_id in required_experts_for_task(task) and action.expert_id not in routed:
routed.append(action.expert_id)
trace.append(
{
"step": obs.step_count,
"action": action.model_dump(exclude_none=True),
"action_label": action_label(action),
"completion_preview": completion[:300],
"reward": round(float(obs.reward), 4),
"consulted_after": list(obs.consulted_experts),
"model_routed_required": list(routed),
}
)
fallback = []
while not obs.done and obs.step_count < obs.max_steps:
act = deterministic_next(obs, task)
obs = env.step(act)
rewards.append(float(obs.reward))
fallback.append(action_label(act))
rows.append(
{
"task": task,
"rag": bool(use_rag),
"action_sequence": [t["action_label"] for t in trace],
"model_routed_required": routed,
"required_experts": required_experts_for_task(task),
"fallback": fallback,
"needed_fallback": bool(fallback),
"policy_reward": round(sum(t["reward"] for t in trace), 4),
"total_reward": round(sum(rewards), 4),
"terminal_score": round(float(obs.terminal_grader_score or 0.0), 4),
"trace": trace,
}
)
eval_dir.mkdir(parents=True, exist_ok=True)
(eval_dir / "evidence.json").write_text(json.dumps(rows, indent=2, default=str), encoding="utf-8")
md = [
"# AutoDataLab++ RL Evidence",
"",
"| Task | RAG | Action sequence | Routed required experts | Needed fallback | Policy reward | Terminal |",
"|---|---:|---|---|---:|---:|---:|",
]
for row in rows:
md.append(
f"| {row['task']} | {row['rag']} | `{' -> '.join(row['action_sequence'])}` | "
f"{', '.join(row['model_routed_required']) or '-'} | {row['needed_fallback']} | "
f"{row['policy_reward']} | {row['terminal_score']} |"
)
(eval_dir / "evidence.md").write_text("\n".join(md), encoding="utf-8")
plt.figure(figsize=(8, 4))
for row in rows:
total = 0.0
curve = []
for t in row["trace"]:
total += float(t["reward"])
curve.append(total)
if curve:
plt.plot(range(1, len(curve) + 1), curve, marker="o", label=f"{row['task']} rag={row['rag']}")
plt.title("RL policy reward on model-controlled steps")
plt.xlabel("policy step")
plt.ylabel("cumulative strict reward")
plt.grid(True, alpha=0.3)
plt.legend()
plt.tight_layout()
plt.savefig(eval_dir / "reward_curve.png", dpi=160)
plt.close()
del tok, model
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return rows
def evidence_score(rows: list[dict[str, Any]]) -> float:
"""Single scalar for quick continue/skip decisions."""
if not rows:
return -999.0
vals = []
for row in rows:
required = set(row.get("required_experts") or [])
routed = set(row.get("model_routed_required") or [])
coverage = len(required & routed) / max(len(required), 1)
no_fallback = 1.0 if not row.get("needed_fallback") else 0.0
policy_reward = float(row.get("policy_reward") or 0.0)
vals.append(coverage + no_fallback + 0.1 * policy_reward)
return sum(vals) / len(vals)
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--method", choices=("grpo", "grpo_rlvr", "ppo"), required=True)
ap.add_argument("--model-id", default="Qwen/Qwen2.5-1.5B-Instruct")
ap.add_argument("--init-adapter", default="", help="optional SFT/DPO adapter directory or HF repo")
ap.add_argument("--run-name", default="")
ap.add_argument("--out-root", type=Path, default=Path("/kaggle/working/cos_1p5b_rl_runs") if Path("/kaggle/working").is_dir() else Path("cos_1p5b_rl_runs"))
ap.add_argument("--hf-token", default=os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN") or "")
ap.add_argument("--no-4bit", action="store_true")
ap.add_argument("--epochs", type=int, default=1)
ap.add_argument(
"--lr",
type=float,
default=5e-6,
help="conservative default; higher LR can collapse a good SFT/DPO adapter to noop",
)
ap.add_argument("--grad-accum", type=int, default=4)
ap.add_argument("--max-length", type=int, default=1024)
ap.add_argument("--variants", type=int, default=2)
ap.add_argument("--max-train-states", type=int, default=0)
ap.add_argument("--log-every", type=int, default=5)
ap.add_argument("--max-grad-norm", type=float, default=0.3)
ap.add_argument("--ppo-clip", type=float, default=0.2)
ap.add_argument(
"--sft-anchor",
type=float,
default=0.2,
help="oracle logprob anchor; keep >0 when continuing from a good SFT/DPO adapter",
)
ap.add_argument("--lora-r", type=int, default=16)
ap.add_argument("--lora-alpha", type=int, default=32)
ap.add_argument("--policy-steps", type=int, default=6)
ap.add_argument("--eval-new-tokens", type=int, default=48)
ap.add_argument("--eval-tasks", default="expert_brief,risk_brief,crisis_brief")
ap.add_argument("--eval-rag-modes", default="false,true", help="comma list: false,true")
ap.add_argument(
"--report-to",
choices=("none", "tensorboard", "wandb"),
default=os.environ.get("REPORT_TO", "tensorboard"),
help="experimental tracking backend for RL loss/reward logs; default tensorboard for judging",
)
ap.add_argument("--wandb-project", default=os.environ.get("WANDB_PROJECT", "autodatalab-plus"))
ap.add_argument("--seed", type=int, default=42)
args = ap.parse_args()
if args.max_train_states == 0:
args.max_train_states = None
set_seed(args.seed)
run_name = args.run_name or f"{args.method}_qwen15b_seed{args.seed}"
out_dir = args.out_root / run_name
out_dir.mkdir(parents=True, exist_ok=True)
(out_dir / "config.json").write_text(json.dumps(vars(args), indent=2, default=str), encoding="utf-8")
print(f"[run] {run_name} -> {out_dir}", flush=True)
before_rows = []
if args.init_adapter:
print("[baseline] evaluating init adapter before RL...", flush=True)
before_rows = evaluate(args, Path(args.init_adapter), out_dir / "eval_before")
print(f"[baseline] evidence_score={evidence_score(before_rows):.4f}", flush=True)
adapter_dir = train_candidate_rl(args, out_dir)
rows = evaluate(args, adapter_dir, out_dir / "eval")
after_score = evidence_score(rows)
before_score = evidence_score(before_rows) if before_rows else None
print("\n=== RL EVIDENCE SUMMARY ===", flush=True)
for row in rows:
print(
f"{row['task']}: {' -> '.join(row['action_sequence'])} | "
f"routed={row['model_routed_required']} | fallback={row['needed_fallback']} | "
f"policy_reward={row['policy_reward']} terminal={row['terminal_score']}",
flush=True,
)
print(f"\n[evidence_score] after={after_score:.4f}", flush=True)
if before_score is not None:
print(f"[evidence_score] before={before_score:.4f}", flush=True)
if after_score + 1e-6 < before_score:
print(
"[decision] SKIP this RL adapter: it regressed versus the init adapter.",
flush=True,
)
else:
print(
"[decision] KEEP this RL adapter: it matched or improved the init adapter.",
flush=True,
)
print(f"\n[adapter] {adapter_dir}", flush=True)
print(f"[eval] {out_dir / 'eval'}", flush=True)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|