scripts/smoke_test2.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=1.8.0",
#     "transformers>=4.45.0",
#     "accelerate>=0.34.0",
#     "scikit-learn",
#     "datasets",
# ]
# ///
"""Second smoke test: verifies ICL bank population + novelty sharpening path
trigger with a stubbed reward that returns 1.0 for a fixed fraction of
completions (regardless of content), so we don't depend on the tiny model
actually reasoning correctly."""

import sys, os, random
sys.path.insert(0, os.path.dirname(__file__))

from datasets import Dataset
from transformers import AutoTokenizer
import transformers.core_model_loading as _cml
_cml.GLOBAL_WORKERS = 1
from trl import GRPOConfig

import xrpo_lib
from xrpo_lib import XRPOTrainer, build_prompt

random.seed(0)


def fake_correctness_reward(completions, answers):
    # Q0 always fails (to force ICL seeding); others succeed 40% of the time
    # so the solved-example bank fills up quickly.
    out = []
    for a in answers:
        if a == "HARD":
            out.append(0.0)
        else:
            out.append(1.0 if random.random() < 0.4 else 0.0)
    return out


xrpo_lib.correctness_reward = fake_correctness_reward  # patch module-level ref used inside the trainer

MODEL = "trl-internal-testing/tiny-Qwen3ForCausalLM"
tokenizer = AutoTokenizer.from_pretrained(MODEL)

questions = [("Q0?", "HARD")] + [(f"Q{i}?", "1") for i in range(1, 6)]
rows = []
for q, a in questions * 6:
    rows.append({"prompt": build_prompt(tokenizer, q), "question": q, "answer": a, "qid": q})
dataset = Dataset.from_list(rows)

def stub_reward(prompts, completions, answer=None, **kwargs):
    return xrpo_lib.correctness_reward(completions, answer)


config = GRPOConfig(
    output_dir="/tmp/xrpo_smoke2",
    per_device_train_batch_size=6,
    num_generations=6,
    gradient_accumulation_steps=1,
    max_steps=25,
    max_completion_length=16,
    report_to=[],
    logging_steps=1,
    beta=0.001,
    push_to_hub=False,
    save_strategy="no",
    dataloader_num_workers=0,
)

trainer = XRPOTrainer(
    model=MODEL,
    reward_funcs=stub_reward,
    train_dataset=dataset,
    args=config,
    use_novelty=True,
    use_icl=True,
    tokenizer_for_icl=tokenizer,
    icl_min_attempts=1,
)

trainer.train()
print("SMOKE TEST 2 OK")
print("icl bank size:", len(trainer.solved_bank))
print("icl injections total:", trainer.icl_injections_total)
print("question_stats:", dict(trainer.question_stats))