File size: 10,781 Bytes
d597642 285599c d597642 ae04e19 d597642 78575eb d597642 78575eb d597642 78575eb d597642 78575eb d597642 78575eb d597642 ae04e19 78575eb ae04e19 d597642 78575eb d597642 78575eb 285599c d597642 78575eb d597642 285599c ae04e19 d597642 ae04e19 d597642 285599c d597642 ae04e19 d597642 ae04e19 d597642 ae04e19 285599c d597642 285599c d597642 | 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 | """OpenSleuth GRPO trainer.
Trains a small Qwen2.5 model with TRL's GRPOTrainer to do in-context program
synthesis — given the public signature of a hidden function plus a handful of
(input, output) probe examples, emit a Python function that reproduces it.
Reward comes from the live OpenSleuth env Space: the agent's code is executed
against the hidden reference under domain-aware fuzzing, and the verifier
returns an `execution_reward - complexity_penalty` score that we hand back to
GRPO as the per-completion reward (plus a tiny formatting shaping reward).
"""
from __future__ import annotations
import argparse
import logging
import os
import sys
import time
import torch
from peft import LoraConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from trl import GRPOConfig, GRPOTrainer
from opensleuth_train import (
EnvClient,
SYSTEM_PROMPT,
build_synthesis_dataset,
discover_functions,
)
from opensleuth_train.reward import format_reward, make_env_reward
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s %(levelname)s %(name)s: %(message)s",
stream=sys.stdout,
)
log = logging.getLogger("opensleuth.train")
def _split_csv(s: str) -> list[str]:
return [x.strip() for x in s.split(",") if x.strip()]
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser()
p.add_argument("--env-url", default=os.environ.get("ENV_URL", "https://anugrah55-opensleuth-env-gemini-cli.hf.space"))
# v0.4 default: switch to Qwen2.5-3B-Instruct for the open-ended task pool.
# The 0.5B baseline saturated easy tasks but couldn't solve the hard /
# Hub-driven ones. 3B + LoRA + 4-bit fits T4-small (16GB).
p.add_argument("--model-name", default=os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-3B-Instruct"))
p.add_argument("--output-dir", default=os.environ.get("OUTPUT_DIR", "/data/opensleuth-grpo"))
p.add_argument(
"--push-to-hub",
default=os.environ.get(
"PUSH_TO_HUB", "anugrah55/opensleuth-qwen2.5-3b-grpo-v2"
),
)
# Task selection / curriculum knobs (v0.4).
p.add_argument(
"--functions",
default=os.environ.get("FUNCTIONS_INCLUDE", ""),
help="Comma-separated subset of task names to train on. Empty = all "
"tasks the env exposes (builtin + Hub).",
)
p.add_argument(
"--difficulty",
default=os.environ.get("DIFFICULTY_FILTER", "all"),
choices=["easy", "medium", "hard", "all"],
help="Curriculum filter: only sample tasks at this difficulty level.",
)
# Difficulty-weighted rollout counts. Replaces the v0.3 single
# n-per-function knob (kept as an optional override).
p.add_argument("--n-easy", type=int, default=int(os.environ.get("N_EASY", "8")))
p.add_argument("--n-medium", type=int, default=int(os.environ.get("N_MEDIUM", "16")))
p.add_argument("--n-hard", type=int, default=int(os.environ.get("N_HARD", "24")))
p.add_argument(
"--n-per-function",
type=int,
default=int(os.environ.get("N_PER_FUNCTION", "0")),
help="If >0, overrides per-difficulty rollout counts with a uniform N.",
)
p.add_argument("--n-probes", type=int, default=int(os.environ.get("N_PROBES", "6")))
# GRPO/optim defaults sized for T4-small (16GB) + Qwen2.5-3B-4bit + LoRA.
p.add_argument("--num-generations", type=int, default=int(os.environ.get("NUM_GENERATIONS", "2")))
p.add_argument("--max-completion-length", type=int, default=int(os.environ.get("MAX_COMPLETION_LENGTH", "384")))
p.add_argument("--max-prompt-length", type=int, default=int(os.environ.get("MAX_PROMPT_LENGTH", "1024")))
p.add_argument("--learning-rate", type=float, default=float(os.environ.get("LEARNING_RATE", "1e-5")))
p.add_argument("--num-train-epochs", type=float, default=float(os.environ.get("NUM_TRAIN_EPOCHS", "1")))
# GRPO requires per_device_train_batch_size to be a multiple of num_generations
# (one prompt is repeated num_generations times, all in the same forward pass).
p.add_argument("--per-device-batch-size", type=int, default=int(os.environ.get("PER_DEVICE_BATCH_SIZE", "2")))
p.add_argument("--gradient-accumulation-steps", type=int, default=int(os.environ.get("GRAD_ACCUM", "4")))
p.add_argument("--no-4bit", action="store_true", default=os.environ.get("NO_4BIT", "0") == "1")
p.add_argument("--seed", type=int, default=int(os.environ.get("SEED", "42")))
return p.parse_args()
def wait_for_env(client: EnvClient, max_wait_s: float = 300.0) -> None:
log.info("waiting for env at %s ...", client.base_url)
start = time.time()
last_err = ""
while time.time() - start < max_wait_s:
try:
h = client.health()
log.info("env healthy: %s", h)
return
except Exception as e: # noqa: BLE001
last_err = str(e)
time.sleep(5)
raise RuntimeError(f"env never became healthy after {max_wait_s}s. Last error: {last_err}")
def main() -> int:
args = parse_args()
log.info("args: %s", vars(args))
client = EnvClient(base_url=args.env_url, timeout=60.0, retries=4)
wait_for_env(client)
include = _split_csv(args.functions) if args.functions else None
difficulty = None if args.difficulty == "all" else args.difficulty
tasks = discover_functions(client, include=include, difficulty=difficulty)
log.info(
"env catalog: %d task(s) after filter (include=%s, difficulty=%s):",
len(tasks), include, difficulty,
)
for t in tasks:
log.info(
" - %-22s difficulty=%-8s source=%s",
t["name"], t.get("difficulty"), t.get("source"),
)
n_per_function_override = args.n_per_function if args.n_per_function > 0 else None
log.info(
"building synthesis dataset (n_easy=%d n_medium=%d n_hard=%d override=%s n_probes=%d)",
args.n_easy, args.n_medium, args.n_hard, n_per_function_override, args.n_probes,
)
dataset = build_synthesis_dataset(
client,
tasks=tasks,
n_per_function=n_per_function_override,
n_easy=args.n_easy,
n_medium=args.n_medium,
n_hard=args.n_hard,
n_probes=args.n_probes,
seed=args.seed,
)
log.info("dataset size: %d rows", len(dataset))
# GRPO with chat-templated prompts: each row needs a "prompt" field, which
# we re-format as a chat message list so the trainer applies the chat
# template under the hood.
def to_chat(row):
return {
"prompt": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": row["prompt"]},
],
"target_function_name": row["target_function_name"],
"row_seed": row["row_seed"],
}
# Drop the human-readable difficulty column from the GRPO-visible map so
# the trainer doesn't try to forward it as a reward-fn kwarg.
drop_cols = [c for c in ("prompt", "difficulty") if c in dataset.column_names]
dataset = dataset.map(to_chat, remove_columns=drop_cols)
# ---- Model + LoRA ----
log.info("loading model %s (4bit=%s)", args.model_name, not args.no_4bit)
bnb_config = None
if not args.no_4bit:
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
)
tokenizer = AutoTokenizer.from_pretrained(args.model_name, trust_remote_code=True)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
peft_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
task_type="CAUSAL_LM",
bias="none",
)
# GRPO requires per_device_train_batch_size to be a multiple of num_generations.
# If the caller didn't pin one explicitly, default to one prompt per device.
per_device_bs = args.per_device_batch_size or args.num_generations
if per_device_bs % args.num_generations != 0:
raise ValueError(
f"per_device_batch_size ({per_device_bs}) must be a multiple of "
f"num_generations ({args.num_generations})."
)
log.info(
"GRPO batching: per_device_batch_size=%d (= %d prompt(s) × %d generations), grad_accum=%d",
per_device_bs, per_device_bs // args.num_generations, args.num_generations,
args.gradient_accumulation_steps,
)
grpo_config = GRPOConfig(
output_dir=args.output_dir,
per_device_train_batch_size=per_device_bs,
gradient_accumulation_steps=args.gradient_accumulation_steps,
learning_rate=args.learning_rate,
num_train_epochs=args.num_train_epochs,
max_prompt_length=args.max_prompt_length,
max_completion_length=args.max_completion_length,
num_generations=args.num_generations,
beta=0.04,
bf16=torch.cuda.is_bf16_supported() if torch.cuda.is_available() else False,
fp16=False,
logging_steps=1,
save_steps=50,
save_total_limit=2,
report_to=[],
seed=args.seed,
push_to_hub=bool(args.push_to_hub) and bool(os.environ.get("HF_TOKEN")),
hub_model_id=args.push_to_hub or None,
hub_strategy="end",
gradient_checkpointing=True,
)
env_reward_fn = make_env_reward(client)
env_reward_fn.__name__ = "env_verifier_reward"
format_reward.__name__ = "format_reward"
# Load the model ourselves so we control quantization + dtype precisely.
# GRPOTrainer in 0.16 takes model objects and passes them through to its
# internal ref-model copy + LoRA wrapping.
log.info("loading base model with quantization=%s", bnb_config is not None)
model_kwargs = {"trust_remote_code": True, "torch_dtype": torch.bfloat16}
if bnb_config is not None:
model_kwargs["quantization_config"] = bnb_config
model = AutoModelForCausalLM.from_pretrained(args.model_name, **model_kwargs)
log.info("instantiating GRPOTrainer")
trainer = GRPOTrainer(
model=model,
reward_funcs=[env_reward_fn, format_reward],
args=grpo_config,
train_dataset=dataset,
peft_config=peft_config,
processing_class=tokenizer,
)
log.info("starting GRPO training")
trainer.train()
log.info("training complete; saving to %s", args.output_dir)
trainer.save_model(args.output_dir)
if grpo_config.push_to_hub:
log.info("pushing to hub: %s", args.push_to_hub)
trainer.push_to_hub()
return 0
if __name__ == "__main__":
sys.exit(main())
|