File size: 2,880 Bytes
274951a | 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 | """Prompt formats shared by RL training, DPO pair generation, and reward eval."""
from __future__ import annotations
CODING_PYTHON_IO_PROMPT = """You are solving a competitive programming problem.
Write a complete Python 3 program that reads from standard input and writes to standard output.
Return only the solution code. Do not include explanations.
Problem:
{problem}"""
# Nemotron's normalized ``prompt`` field already contains this instruction
# wrapper. Keeping it inside our own stricter code-only wrapper creates
# conflicting instructions and makes Qwen3.5 emit long step-by-step comments
# inside an unfinished code fence even when chat-template thinking is disabled.
# Match the complete prefix exactly so ordinary problem statements are never
# rewritten merely because they mention similar words.
NEMOTRON_LEGACY_PROMPT_PREFIX = """You are a helpful and harmless assistant. You should think step-by-step before responding to the instruction below.
Please use python programming language only.
You must use ```python for just the final solution code block with the following format:
```python
# Your code here
```
"""
def normalize_coding_problem(problem: str) -> str:
problem = problem.strip()
if problem.startswith(NEMOTRON_LEGACY_PROMPT_PREFIX):
problem = problem.removeprefix(NEMOTRON_LEGACY_PROMPT_PREFIX).strip()
return problem
def format_coding_prompt(problem: str, *, format_name: str = "python_io") -> str:
problem = normalize_coding_problem(problem)
if format_name == "raw":
return problem
if format_name == "python_io":
return CODING_PYTHON_IO_PROMPT.format(problem=problem)
raise ValueError(f"unknown coding prompt format {format_name!r}")
def render_coding_prompt(tokenizer, problem: str, *, format_name: str = "python_io") -> str:
"""Render the exact prompt string the policy is trained/evaluated on.
Chat-tuned models (all Qwen3.5 sizes) must see their chat template — feeding
them raw text puts them in continuation mode and they ramble instead of
answering. ``enable_thinking=False`` pins non-thinking mode across model
sizes whose templates default differently; templates without that variable
ignore it. Tokenizers without a chat template fall back to the raw format.
Every consumer of coding prompts (GRPO rollouts, the per-checkpoint reward
callback, DPO pair generation, prompt-length filtering) must go through this
one function so the token stream is identical everywhere.
"""
user_text = format_coding_prompt(problem, format_name=format_name)
if getattr(tokenizer, "chat_template", None):
return tokenizer.apply_chat_template(
[{"role": "user", "content": user_text}],
tokenize=False,
add_generation_prompt=True,
enable_thinking=False,
)
return user_text
|