File size: 9,025 Bytes
51dc497
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import ast
import json
import multiprocessing as mp
import re
from dataclasses import dataclass, field
from typing import Any, Optional

import trackio
from datasets import load_dataset
from transformers import AutoTokenizer, HfArgumentParser, TrainerCallback
from trl import GRPOConfig, GRPOTrainer


SYSTEM_PROMPT = (
    "You are a Python code generator. Return only Python code. "
    "Write a correct solution function for the task. Include a concise but proper docstring on the main function."
)


@dataclass
class ScriptArgs:
    model_name_or_path: str = field(default="Qwen/Qwen3-1.7B-Base")
    dataset_name: str = field(default="google-research-datasets/mbpp")
    dataset_config: str = field(default="sanitized")
    train_split: str = field(default="train")
    eval_split: str = field(default="validation")
    max_train_samples: Optional[int] = field(default=300)
    max_eval_samples: Optional[int] = field(default=64)
    attn_implementation: str = field(default="sdpa")
    reward_timeout: int = field(default=4)
    push_to_hub: bool = field(default=True)
    hub_model_id: str = field(default="AbhilekhMeda/qwen3-1.7b-grpo-python-mbpp")
    run_name: str = field(default="grpo_qwen3_1p7b_mbpp_exec_reward")
    project: str = field(default="grpo-qwen3-python-code")
    trackio_space_id: str = field(default="AbhilekhMeda/mlintern-grpoqwen")


def _strip_code_fence(text: str) -> str:
    text = text.strip()
    match = re.search(r"```(?:python)?\n(.*?)```", text, re.DOTALL | re.IGNORECASE)
    return match.group(1).strip() if match else text


def _extract_function_name_from_tests(test_list: list[str]) -> Optional[str]:
    for test in test_list:
        m = re.search(r"assert\s+([A-Za-z_][A-Za-z0-9_]*)\s*\(", test)
        if m:
            return m.group(1)
    return None


def _build_prompt(example: dict[str, Any]) -> dict[str, Any]:
    prompt_text = (
        f"Task:\n{example['prompt']}\n\n"
        "Requirements:\n"
        "1. Return only Python code.\n"
        "2. Implement the requested function.\n"
        "3. Include a proper docstring on the main function.\n"
        "4. Do not print example usage.\n"
    )
    return {
        "prompt": [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": prompt_text},
        ],
        "task_id": example["task_id"],
        "raw_prompt": example["prompt"],
        "test_imports": example["test_imports"],
        "test_list": example["test_list"],
        "reference_code": example["code"],
    }


def _docstring_ok(code: str, fn_name: Optional[str]) -> bool:
    if not fn_name:
        return False
    try:
        tree = ast.parse(code)
        for node in tree.body:
            if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)) and node.name == fn_name:
                doc = ast.get_docstring(node)
                return bool(doc and len(doc.strip()) >= 10)
    except Exception:
        return False
    return False


def _runner(queue, code: str, imports: list[str], tests: list[str]):
    glb: dict[str, Any] = {}
    try:
        exec(compile(code, "<candidate>", "exec"), glb, glb)
        if imports:
            exec("\n".join(imports), glb, glb)
        for test in tests:
            exec(test, glb, glb)
        queue.put({"ran": True, "passed": True, "error": ""})
    except Exception as e:
        queue.put({"ran": True, "passed": False, "error": repr(e)})


def _exec_in_subprocess(code: str, test_imports: list[str], tests: list[str], timeout: int) -> dict[str, Any]:
    ctx = mp.get_context("spawn")
    queue = ctx.Queue()
    proc = ctx.Process(target=_runner, args=(queue, code, test_imports, tests))
    proc.start()
    proc.join(timeout)
    if proc.is_alive():
        proc.terminate()
        proc.join(1)
        return {"ran": False, "passed": False, "error": "timeout"}
    return queue.get() if not queue.empty() else {"ran": False, "passed": False, "error": "no_result"}


def execution_reward(completions, test_imports, test_list, log_extra=None, log_metric=None, **kwargs):
    rewards = []
    run_rates, pass_rates, doc_rates, errors = [], [], [], []
    for completion, imports, tests in zip(completions, test_imports, test_list):
        content = completion[0]["content"] if isinstance(completion, list) else completion
        code = _strip_code_fence(content)
        fn_name = _extract_function_name_from_tests(tests)
        result = _exec_in_subprocess(code, imports, tests, timeout=4)
        ran = 1.0 if result["ran"] and result["error"] != "timeout" else 0.0
        passed = 1.0 if result["passed"] else 0.0
        doc = 1.0 if _docstring_ok(code, fn_name) else 0.0
        rewards.append(0.25 * ran + 0.6 * passed + 0.15 * doc)
        run_rates.append(ran)
        pass_rates.append(passed)
        doc_rates.append(doc)
        errors.append(result["error"][:120])
    if log_extra:
        log_extra("exec_error", errors)
        log_extra("passed_tests", [str(x) for x in pass_rates])
        log_extra("docstring_ok", [str(x) for x in doc_rates])
    if log_metric and rewards:
        n = float(len(rewards))
        log_metric("exec_run_rate", sum(run_rates) / n)
        log_metric("exec_pass_rate", sum(pass_rates) / n)
        log_metric("docstring_rate", sum(doc_rates) / n)
    return [float(r) for r in rewards]


class AlertCallback(TrainerCallback):
    def on_log(self, args, state, control, logs=None, **kwargs):
        if not logs:
            return
        if logs.get("loss") is not None and logs["loss"] > 2.0:
            trackio.alert("high_loss", f"loss={logs['loss']:.4f} at step {state.global_step} — try lr x0.5 if it persists", level="WARN")
        if logs.get("reward") is not None and logs["reward"] > 0.8:
            trackio.alert("strong_reward", f"reward={logs['reward']:.4f} at step {state.global_step} — keep current config and refine around lr", level="INFO")
        if logs.get("completions/clipped_ratio") is not None and logs["completions/clipped_ratio"] > 0.3:
            trackio.alert("high_clipping", f"clipped_ratio={logs['completions/clipped_ratio']:.4f} at step {state.global_step} — increase max_completion_length if many outputs are truncated", level="WARN")


def main():
    parser = HfArgumentParser((ScriptArgs, GRPOConfig))
    script_args, training_args = parser.parse_args_into_dataclasses()

    dataset = load_dataset(script_args.dataset_name, script_args.dataset_config)
    train_dataset = dataset[script_args.train_split]
    eval_dataset = dataset[script_args.eval_split]
    if script_args.max_train_samples:
        train_dataset = train_dataset.select(range(min(script_args.max_train_samples, len(train_dataset))))
    if script_args.max_eval_samples:
        eval_dataset = eval_dataset.select(range(min(script_args.max_eval_samples, len(eval_dataset))))

    train_dataset = train_dataset.map(_build_prompt, remove_columns=train_dataset.column_names)
    eval_dataset = eval_dataset.map(_build_prompt, remove_columns=eval_dataset.column_names)

    tokenizer = AutoTokenizer.from_pretrained(script_args.model_name_or_path, padding_side="left")
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    training_args.report_to = "trackio"
    training_args.run_name = script_args.run_name
    training_args.project = script_args.project
    training_args.trackio_space_id = script_args.trackio_space_id
    training_args.push_to_hub = script_args.push_to_hub
    training_args.hub_model_id = script_args.hub_model_id
    training_args.disable_tqdm = True
    training_args.logging_strategy = "steps"
    training_args.logging_first_step = True
    training_args.save_strategy = "steps"
    training_args.eval_strategy = "steps"
    training_args.remove_unused_columns = False
    training_args.model_init_kwargs = {
        "attn_implementation": script_args.attn_implementation,
        "torch_dtype": "bfloat16",
    }
    training_args.chat_template_kwargs = {"enable_thinking": False}

    trainer = GRPOTrainer(
        model=script_args.model_name_or_path,
        args=training_args,
        processing_class=tokenizer,
        reward_funcs=[execution_reward],
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        callbacks=[AlertCallback()],
    )

    trackio.alert("run_start", f"Starting GRPO code run with train_samples={len(train_dataset)} eval_samples={len(eval_dataset)} lr={training_args.learning_rate}", level="INFO")
    trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
    metrics = trainer.evaluate()
    trainer.log_metrics("eval", metrics)
    trainer.save_metrics("eval", metrics)
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(commit_message="End of GRPO training")
    trackio.alert("run_complete", f"Training complete at step {trainer.state.global_step} with eval metrics: {json.dumps(metrics, default=str)}", level="INFO")


if __name__ == "__main__":
    main()