File size: 17,055 Bytes
ff4becd | 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 | """Qwen-3.5-27B QLoRA fine-tune driver for the deferred Family-7 slot.
Plan reference: Phase 3.1 (R1 Q1.3, R2 Q2.4, R3 W3.4 saturation, R5 W5.1;
author A2). T3 and T6 saturate near 100% MAPE for every zero-shot LLM
in the panel; they are the cleanest demonstration target for whether
MacroLens supports supervised LLM training. This driver populates the
single deferred Family-7 column in Tables 6-10.
Design choices (orthodox interpretation of the existing
:mod:`methods.llm_finetune` infrastructure):
* **Per-task adapters.** :class:`methods.LLMFineTuned` fixes ``self.task``
at construction and dispatches on it; we therefore train TWO adapters,
one for T3 and one for T6, rather than one bundled multi-task adapter.
Both tasks share the same JSON output schema (11 canonical XBRL
fields); the per-task split keeps the prompt format precisely matched
to each task. This is the simplest configuration that uses the
existing class as-is.
* **Cross-task transfer.** The T3 adapter is evaluated zero-shot on
T1 / T2 / T4 / T5 / T7 as a catastrophic-forgetting check: if a single
task's QLoRA pass leaves the model's competence on other tasks intact,
the result-table column can be populated end-to-end; otherwise the
cross-task entries become Family-7 / not-applicable.
* **Library defaults.** ``LLMFineTunedConfig`` ships with
``lora_r=16, lora_alpha=32, epochs=3, learning_rate=2e-4`` -- this
is what the panel-FT recipe was originally pre-registered to use. We
do not vary any of those four numbers in this driver, per the
no-tuning rule.
* **Single seed.** ``panel.PRIMARY_SEED = 42`` end-to-end.
Per-launch authorisation: this is a multi-hour GPU run (4 x A100-40GB,
GPU IDs 4-7 per project memory). The user must authorise the launch.
"""
from __future__ import annotations
import argparse
import json
import logging
import pickle
import time
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any
import numpy as np
logger = logging.getLogger(__name__)
# Tasks evaluated by the trained adapters. Native-task targets are
# the rows that directly populate the Family-7 column; cross-task
# targets test for catastrophic forgetting.
_NATIVE_TASKS: tuple[str, ...] = ("T3", "T6")
_CROSS_TASKS: tuple[str, ...] = ("T1", "T2", "T4", "T5", "T7")
# Primary metric per task (mirrors gen_tables.py conventions).
_PRIMARY_METRIC: dict[str, str] = {
"T1": "mse",
"T2": "medape",
"T3": "mape",
"T4": "mae",
"T5": "medape",
"T6": "mape",
"T7": "mape",
}
_CLUSTER_KEY: dict[str, str] = {
"T1": "ticker",
"T2": "ticker",
"T3": "ticker",
"T4": "scenario_id",
"T5": "ticker",
"T6": "ticker",
"T7": "address",
}
@dataclass
class _EvalCell:
adapter_task: str
eval_task: str
is_native: bool
seed: int
n_test: int
primary_metric: str
value: float
ci_lo: float
ci_hi: float
fit_sec: float | None
predict_sec: float
def _cluster_keys(task: str, meta_test: Any) -> Any:
key = _CLUSTER_KEY[task]
if hasattr(meta_test, "columns") and key in meta_test.columns:
return meta_test[key].values
if hasattr(meta_test, "get"):
keys = meta_test.get(key)
if keys is not None:
return np.asarray(keys)
return None
def _save_predictions(
*,
pred_dir: Path,
method_id: str,
task: str,
seed: int,
granularity: str,
y_pred: Any,
y_test: Any,
meta_test: Any,
extra_tag: str | None = None,
) -> Path:
pred_dir.mkdir(parents=True, exist_ok=True)
tag = f"{method_id}_{task}_seed{seed}"
if extra_tag:
tag += f"_{extra_tag}"
out_path = pred_dir / f"{tag}.pkl"
tmp = out_path.with_suffix(".pkl.tmp")
with open(tmp, "wb") as f:
pickle.dump({
"method_id": method_id,
"task": task,
"seed": seed,
"granularity": granularity,
"y_pred": y_pred,
"y_test": y_test,
"meta_test": meta_test,
"timestamp": time.strftime("%Y-%m-%dT%H:%M:%S"),
}, f)
tmp.replace(out_path)
return out_path
def _engine_for_adapter(
*, base_hf_id: str, adapter_path: Path, base_url: str, api_key: str,
) -> Any:
"""Build an OpenAI-compatible engine targeting the vLLM-served adapter.
The runner exposes the adapter as a LoRA module via:
vllm serve <base_hf_id> --enable-lora \\
--lora-modules adapter_qwen35_t3=<adapter_path>
so ``model_id`` resolves to the LoRA name, not the base HF id.
"""
from projects.agent_builder.scripts.whatif_bench.methods._openai_engine import (
OpenAIEngine,
)
return OpenAIEngine(
base_url=base_url, api_key=api_key,
model_id=str(adapter_path.name), # vLLM LoRA module id is the dir name
)
def train_adapter(
*,
task: str,
base_model: str = "qwen35",
granularity: str = "daily",
seed: int = 42,
adapter_dir: Path,
) -> tuple[Path, float]:
"""Train a single-task QLoRA adapter using :class:`LLMFineTuned`.
Returns ``(adapter_path, fit_sec)``.
"""
from projects.agent_builder.scripts.whatif_bench import macrolens as ml
from projects.agent_builder.scripts.whatif_bench.methods.llm_finetune import (
LLMFineTuned,
)
from projects.agent_builder.scripts.whatif_bench.methods._config import (
LLMFineTunedConfig,
)
train = ml.load(task, "train", granularity=granularity)
cfg = LLMFineTunedConfig() # library-default lora_r / lora_alpha / epochs / lr
model = LLMFineTuned(task=task, config=cfg, base_model=base_model)
t0 = time.perf_counter()
model.fit(train.X, train.y, seed=seed)
fit_sec = time.perf_counter() - t0
adapter_dir.mkdir(parents=True, exist_ok=True)
out_path = adapter_dir / f"qwen35_qlora_{task.lower()}"
model.save(out_path)
logger.info("trained %s adapter -> %s (fit %.1fs)", task, out_path, fit_sec)
return out_path, fit_sec
# [REVERTED 2026-05-19] An earlier in-session draft of
# ``train_multitask_adapter`` was inserted here but used
# ``device_map="auto"`` + bnb-4bit on Llama-4 Scout MoE, which
# RESEARCH_PLAN.md §5.1 + IMPLEMENTATION_PLAN.md §F7 explicitly call
# out as the documented "MoE-on-bitsandbytes complexity" failure mode
# (Scout needs ZeRO-2 across 4 GPUs, not naive auto-placement). The
# draft was reverted so the canonical sources (Llama-4 Scout HF card,
# Meta torchtune SFT example, HF PEFT MoE docs, TRL response-only-loss
# docs, DeepSpeed ZeRO-2 config) can be read end-to-end first and a
# verified recipe written rather than improvised.
def evaluate_with_adapter(
*,
adapter_path: Path,
adapter_task: str,
eval_task: str,
base_hf_id: str,
base_url: str,
api_key: str,
granularity: str,
seed: int,
pred_dir: Path,
fit_sec: float | None,
) -> _EvalCell:
"""Evaluate an adapter on ``eval_task``.
For native eval (``eval_task == adapter_task``) the predict path is
the task-native predict path on :class:`LLMFineTuned`. For cross-task
eval we still use :class:`LLMFineTuned` so the prompt formatting is
consistent with the panel's other LLM-FT cells; the adapter is
loaded fresh, then ``model.task`` is overridden to ``eval_task`` so
the right per-task predict path runs.
"""
from projects.agent_builder.scripts.whatif_bench import macrolens as ml
from projects.agent_builder.scripts.whatif_bench.methods.llm_finetune import (
LLMFineTuned,
)
engine = _engine_for_adapter(
base_hf_id=base_hf_id, adapter_path=adapter_path,
base_url=base_url, api_key=api_key,
)
test = ml.load(eval_task, "test", granularity=granularity)
model = LLMFineTuned.load(adapter_path)
# ``LLMFineTuned.load`` reconstructs at the trained-task. Force the
# task for cross-eval; the trained QLoRA adapter is unchanged.
model.task = eval_task
model.engine = engine
t1 = time.perf_counter()
y_pred = model.predict(test.X)
predict_sec = time.perf_counter() - t1
_save_predictions(
pred_dir=pred_dir,
method_id="llm_finetuned_qwen35",
task=eval_task,
seed=seed,
granularity=granularity,
y_pred=y_pred,
y_test=test.y,
meta_test=test.meta,
extra_tag=(
None if eval_task == adapter_task
else f"transfer_from_{adapter_task}"
),
)
metrics = ml.score(
eval_task, test.y, y_pred,
cluster_keys=_cluster_keys(eval_task, test.meta),
resample="cluster", n_boot="adaptive", seed=seed,
)
primary = _PRIMARY_METRIC[eval_task]
mv = metrics[primary]
value = float("nan") if mv.value is None else float(mv.value)
ci_lo = float("nan") if mv.ci_lo is None else float(mv.ci_lo)
ci_hi = float("nan") if mv.ci_hi is None else float(mv.ci_hi)
return _EvalCell(
adapter_task=adapter_task,
eval_task=eval_task,
is_native=(eval_task == adapter_task),
seed=seed,
n_test=int(len(test.y)) if hasattr(test.y, "__len__") else -1,
primary_metric=primary,
value=value,
ci_lo=ci_lo,
ci_hi=ci_hi,
fit_sec=fit_sec if eval_task == adapter_task else None,
predict_sec=predict_sec,
)
def run_pipeline(
*,
base_url: str | None,
api_key: str = "EMPTY",
base_model: str = "qwen35",
granularity: str = "daily",
seed: int | None = None,
adapter_dir: Path | None = None,
pred_dir: Path | None = None,
eval_native_only: bool = False,
train_only: bool = False,
multitask: bool = False,
) -> dict[str, Any]:
"""End-to-end pipeline: train (T3, T6) adapters then evaluate.
Two-pass: (i) train each native-task adapter; (ii) evaluate each
adapter on its native task plus the cross-task panel (T3 adapter
only, to keep the GPU budget bounded).
"""
from projects.agent_builder.scripts.whatif_bench.experiments import panel
from projects.agent_builder.scripts.whatif_bench.methods.llm_finetune import (
_BASE_MODEL_ID,
)
seed = seed if seed is not None else panel.PRIMARY_SEED
adapter_dir = adapter_dir or Path(__file__).resolve().parents[1] / "adapters"
pred_dir = pred_dir or Path(__file__).resolve().parents[1] / "predictions"
base_hf_id = _BASE_MODEL_ID.get(base_model)
if base_hf_id is None:
raise ValueError(f"unknown base_model {base_model!r}; "
f"expected one of {sorted(_BASE_MODEL_ID)}")
if multitask:
raise NotImplementedError(
"multitask=True was reverted on 2026-05-19 pending re-read of "
"Llama-4 Scout / DeepSpeed ZeRO-2 / PEFT-MoE primary sources. "
"See header comment near the reverted train_multitask_adapter "
"block."
)
# (i) Train adapters.
adapters: dict[str, tuple[Path, float]] = {}
multitask_pair_counts: dict[str, int] = {}
for task in _NATIVE_TASKS:
adapter_path, fit_sec = train_adapter(
task=task, base_model=base_model, granularity=granularity,
seed=seed, adapter_dir=adapter_dir,
)
adapters[task] = (adapter_path, fit_sec)
# Train-only short-circuit: skip the eval phase entirely. Used to
# separate the long-running QLoRA training step from the eval step,
# which requires a separately-orchestrated vLLM serve endpoint.
if train_only:
return {
"probe": "llm_finetune_scout_multitask" if multitask else "llm_finetune_qwen35",
"base_model": base_model,
"base_hf_id": base_hf_id,
"granularity": granularity,
"seed": seed,
"multitask": multitask,
"multitask_pair_counts": multitask_pair_counts,
"adapters": {
task: {"path": str(path), "fit_sec": fs}
for task, (path, fs) in adapters.items()
},
"cells": [],
"train_only": True,
}
if base_url is None:
raise ValueError(
"run_pipeline: --base-url required when --train-only is not set "
"(eval needs a vLLM serve endpoint serving the trained adapters)."
)
# (ii) Evaluate. Native-task eval per adapter; cross-task eval uses
# the T3 adapter only (T3 train is ~7x the size of T6 train and
# produces the more general checkpoint).
cells: list[_EvalCell] = []
for adapter_task, (adapter_path, fit_sec) in adapters.items():
cell = evaluate_with_adapter(
adapter_path=adapter_path, adapter_task=adapter_task,
eval_task=adapter_task, base_hf_id=base_hf_id,
base_url=base_url, api_key=api_key, granularity=granularity,
seed=seed, pred_dir=pred_dir, fit_sec=fit_sec,
)
cells.append(cell)
if not eval_native_only:
t3_path, _ = adapters["T3"]
for cross in _CROSS_TASKS:
try:
cell = evaluate_with_adapter(
adapter_path=t3_path, adapter_task="T3",
eval_task=cross, base_hf_id=base_hf_id,
base_url=base_url, api_key=api_key,
granularity=granularity, seed=seed,
pred_dir=pred_dir, fit_sec=None,
)
cells.append(cell)
except Exception as exc:
logger.exception("cross-task eval %s failed: %s", cross, exc)
cells.append(_EvalCell(
adapter_task="T3", eval_task=cross, is_native=False,
seed=seed, n_test=-1,
primary_metric=_PRIMARY_METRIC[cross],
value=float("nan"), ci_lo=float("nan"), ci_hi=float("nan"),
fit_sec=None, predict_sec=float("nan"),
))
return {
"probe": "llm_finetune_qwen35",
"base_model": base_model,
"base_hf_id": base_hf_id,
"base_url": base_url,
"granularity": granularity,
"seed": seed,
"adapters": {
task: {
"path": str(path),
"fit_sec": fs,
}
for task, (path, fs) in adapters.items()
},
"cells": [asdict(c) for c in cells],
}
def _default_probe_dir() -> Path:
# Probe outputs live under experiments/ (experiment artifacts),
# never under data_small_caps/ (raw + derived benchmark data).
return Path(__file__).resolve().parents[1] / "probes_output"
def main() -> int:
parser = argparse.ArgumentParser(
description="Qwen-3.5-27B QLoRA fine-tune driver (Phase 3.1).",
)
parser.add_argument("--base-url", default=None,
help="vLLM OpenAI-compatible endpoint (e.g., http://localhost:8004/v1). "
"Required unless --train-only is set.")
parser.add_argument("--api-key", default="EMPTY")
parser.add_argument("--base-model", default="qwen35",
choices=["llama_scout", "gemma4", "qwen35"])
parser.add_argument("--granularity", default="daily")
parser.add_argument("--seed", type=int, default=None)
parser.add_argument("--adapter-dir", type=Path, default=None)
parser.add_argument("--pred-dir", type=Path, default=None)
parser.add_argument("--eval-native-only", action="store_true",
help="Skip cross-task transfer eval (T1/T2/T4/T5/T7).")
parser.add_argument("--train-only", action="store_true",
help="Train adapters and exit; skip the eval phase "
"(which requires a vLLM serve endpoint).")
parser.add_argument("--multitask", action="store_true",
help="Train ONE adapter on a pooled corpus over all "
"7 task train splits (T1..T7). Default is the "
"per-task design (T3+T6 only with cross-task "
"eval). Recommended for Phase 3.1 Family-7.")
parser.add_argument("--output", type=Path, default=None,
help="Path to the summary JSON report.")
args = parser.parse_args()
logging.basicConfig(
level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s",
)
report = run_pipeline(
base_url=args.base_url, api_key=args.api_key,
base_model=args.base_model, granularity=args.granularity,
seed=args.seed, adapter_dir=args.adapter_dir,
pred_dir=args.pred_dir, eval_native_only=args.eval_native_only,
train_only=args.train_only, multitask=args.multitask,
)
out_path = args.output or _default_probe_dir() / "llm_finetune_qwen35.json"
out_path.parent.mkdir(parents=True, exist_ok=True)
out_path.write_text(json.dumps(report, indent=2, default=str))
logger.info("fine-tune report written to %s", out_path)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|