touristgpt/TouristGPT
TouristGPT is a 7.6 billion parameter instruction-tuned causal language model developed by touristgpt, based on the Qwen2.5-Coder-7B-Instruct architecture. It is specifically fine-tuned over 10 epochs for C++ competitive programming, excelling at generating complete, submission-ready C++17 solutions. The model achieves a 22.0% full pass rate on internal Codeforces benchmarks with zero compile errors, making it highly specialized for automated competitive programming tasks.
Loading preview...
TouristGPT: Specialized for C++ Competitive Programming
TouristGPT is a 7.6 billion parameter model derived from Qwen/Qwen2.5-Coder-7B-Instruct, meticulously fine-tuned for C++ competitive programming. The model underwent 10 epochs of training using a length-sorted Curriculum Sampler, with Epoch 7 selected for its stability and performance.
Key Capabilities
- C++ Code Generation: Generates complete, submission-ready C++17 solutions, including necessary headers,
main(), andsolve()functions. - Reasoning Block: Incorporates a
<think>...</think>block for natural language reasoning, constraint checking, and sample validation before code generation. - High Reliability: Achieved a 22.0% Full Pass Rate on an internal blind test set of 50 Codeforces problems, meaning generated solutions compiled and passed all sample test cases zero-shot.
- Zero Compile Errors: Demonstrated a 100% syntactical compilation success rate on generated solutions in internal benchmarks.
- Strict Output Format: Designed to adhere to a precise output format, providing only a reasoning block and a code block.
Usage Guidelines
TouristGPT utilizes the standard ChatML format and requires a specific system prompt instructing it to think step-by-step. Users must also adhere to a strict user prompt format, including problem statements, time/memory limits, input/output descriptions, and examples, to match the model's training curriculum perfectly. This structured input ensures optimal performance for competitive programming challenges.