MicroCoder - A large model training optimization framework developed by Microsoft in collaboration with Cambridge and others.

What is MicroCoder?

MicroCoder is a code model training optimization project launched by Microsoft Research Asia in collaboration with the University of Cambridge and Princeton University. It addresses the training bottlenecks of next-generation models such as Qwen3 by comprehensively upgrading four dimensions: algorithm, data, evaluation, and experience. The core framework includes: the MicroCoder-GRPO algorithm (conditional truncation mask, diversity temperature selection, and removal of KL divergence), a dataset of 13K+ real-world competition problems, a highly fault-tolerant evaluation framework, and 34 training insights across seven dimensions, providing a complete solution for reinforcement learning of modern code models.

MicroCoder's main functions

• Algorithm optimizationThe MicroCoder-GRPO algorithm optimizes the reinforcement learning training process of modern code models through three key improvements.
• Data buildingMicroCoder-Dataset provides over 13,000 rigorously selected real-world competitive programming problems for model training.
• Enhanced assessmentMicroCoder-Evaluator employs a multi-method backchain mechanism to improve the accuracy of code evaluation and training efficiency.
• Summary of experienceMicroCoder-Insights has summarized 34 training experiences covering seven dimensions based on more than 30 controlled experiments.

MicroCoder's technical principles

• Algorithm PrinciplesThe MicroCoder-GRPO algorithm optimizes for the significantly different training dynamics of next-generation code models compared to older models. It introduces a conditional truncation masking mechanism, masking only outputs that simultaneously meet four conditions: reaching maximum length, correct answer, no tail repetition, and random probability sampling. This unlocks the model's long output potential while avoiding training problems associated with full masking strategies. The algorithm proposes a method to dynamically determine the training temperature based on the initial output diversity, finding that a phased temperature strategy (lower temperature initially, higher temperature later) outperforms a fixed temperature throughout. Furthermore, the algorithm sets the KL divergence weight to zero and employs a higher pruning ratio, eliminating the continuous suppression of output diversity by KL divergence and enabling long-term stable performance improvements.
• Data PrinciplesThe MicroCoder-Dataset was built using a four-stage data processing pipeline, including collecting real competition problems from multiple platforms, standardizing the format and denoising, implementing soft and hard constraints and adaptive difficulty filtering, and manual verification. The core innovation of the dataset lies in its five-dimensional difficulty assessment matrix. This matrix references Bloom's educational objective classification method and code complexity metrics. Each problem was scored independently three times by an LLM (Local Modeling Module) and then weighted and averaged. The model's actual pass rate was used as a benchmark for calibration, ultimately increasing the proportion of difficult problems to over 50%, ensuring that the training data matches the capabilities of the next-generation model.
• Evaluation PrinciplesMicroCoder-Evaluator addresses the issue of numerous misjudgments caused by the strict and precise matching strategy of the original LiveCodeBench evaluator by designing a backtracking chain comprehensive verification mechanism consisting of 6 to 7 methods. This mechanism sequentially attempts different comparison strategies, including automatic type conversion for formats such as lists, tuples, strings, and sets, performing floating-point approximation comparisons, and preprocessing such as multi-line splitting and whitespace normalization. It automatically switches to the next method after a single method fails, maintaining high fault tolerance throughout the process. This improves evaluation accuracy by approximately 25% and increases the execution speed of training steps by approximately 40% through optimized parallel processing strategies.

How to use MicroCoder

• Environmental preparationClone the MicroCoder GitHub repository to your local environment and install the relevant dependencies.
• Data AcquisitionDownload the MicroCoder-Dataset dataset or build your own high-quality programming problem dataset following the documentation.
• Algorithm ConfigurationSelect hyperparameters of the MicroCoder-GRPO algorithm based on model characteristics, including temperature strategy and mask ratio.
• Evaluation settingsDeploy MicroCoder-Evaluator to replace the standard evaluator to ensure accurate feedback signals during training.
• Model training: Start the reinforcement learning training process with the configured algorithm, dataset and evaluator.
• Application of experienceRefer to the 34 training insights in MicroCoder-Insights to optimize various settings during the training process.
• Effect verificationTest the performance of the trained model on code benchmarks such as LiveCodeBench.

Key information and usage instructions for MicroCoder

• Project BackgroundMicroCoder, jointly launched by Microsoft Research Asia, Cambridge University, and Princeton University, addresses the issue of traditional methods failing during the training of next-generation code models such as Qwen3, and features a comprehensive upgrade across four dimensions: algorithm, data, evaluation, and experience.
• Core componentsIncludes the MicroCoder-GRPO algorithm (three key improvements), a dataset of 13K+ real-world competition problems, a highly fault-tolerant evaluation framework, and 34 training insights across seven dimensions.
• Technical HighlightsConditional truncation mask balances long output and stability; diversity-driven temperature selection maintains training diversity; KL divergence removal achieves continuous improvement; and a five-dimensional difficulty evaluation matrix ensures data quality.
• Performance improvementCompared to the DeepCoder dataset, the training performance is improved by 3 times, the evaluation accuracy is improved by 25%, and the training speed is improved by 40%.

MicroCoder's core advantages

• intergenerational adaptation advantagesMicroCoder is the first to identify and address the fundamental differences in training dynamics between new and old code models. It is specifically optimized to address the characteristics of continuously increasing output length of new-generation models such as Qwen3 and insufficient difficulty of traditional datasets, breaking the inherent perception of directly transferring mathematical reasoning training experience.
• Algorithm Innovation AdvantagesMicroCoder-GRPO achieves continuous performance improvement through three synergistic effects: it uses conditional truncation masks to finely control the potential for long outputs, employs diversity-driven temperature selection to maintain training stability, and removes KL divergence to eliminate performance bottlenecks. These improvements work together to achieve continuous performance improvement rather than an initial rise followed by a fall.
• Data quality advantagesMicroCoder-Dataset contains 13,000+ real competition problems that are not generated by LLM. Through a five-dimensional difficulty assessment matrix and a prediction-calibration-selection mechanism, the proportion of difficult problems is increased to more than 50%, ensuring that the training data matches the capabilities of the next-generation model. The performance gain is up to 3 times that of DeepCoder under the same number of training steps.
• Accurate assessment advantagesMicroCoder-Evaluator uses 6-7 backchain methods to replace strict exact matching, supports type conversion, floating-point approximation and other high-fault-tolerant verification, improves accuracy by 25% and training speed by 40%, and provides more reliable feedback signals for reinforcement learning.
• Advantages of accumulated experienceThe project summarizes 34 training insights through 30+ controlled experimental systems, covering seven dimensions: evaluator, temperature, data, context, mask, batch size, and KL divergence. It is one of the most complete knowledge accumulations in the field of post-training of large-scale code models to date.

MicroCoder's project address

• GitHub repositoryhttps://github.com/ZongqianLi/MicroCoder

MicroCoder's Competitive Product Comparison

Comparison Dimensions	MicroCoder	DeepCoder	Standard GRPO/DAPO
Core positioning	Next-generation code model training and optimization framework	Mainstream code training datasets	Traditional reinforcement learning algorithms
Data difficulty	13,000+ real competition problems, with over 50% being difficult.	The problem was too easy for the new model, and training yielded almost no improvement.	Relying on external datasets does not solve the problem of data difficulty.
Algorithm performance	Continuous and stable improvement, with no performance bottlenecks.	No proprietary algorithm	The new model exhibits issues such as limited output length and a pattern of initial rise followed by fall.
Key Innovation	Three improvements: conditional masking, dynamic temperature, and removal of KL divergence.	none	Basic GRPO strategy

Application scenarios of MicroCoder

• Large code model developmentMicroCoder is suitable for reinforcement learning post-training of next-generation inference code models, helping R&D teams overcome the training bottlenecks of traditional methods on advanced models such as Qwen3 and achieve more efficient performance improvements.
• Competition Programming Model OptimizationFor scenarios such as algorithm competitions and programming competitions, the MicroCoder-Dataset provides high-quality real-world competition problem data to train models to master the reasoning and coding abilities for complex algorithm problems.
• Enterprise Code Assistant DevelopmentEnterprises can build internal code generation assistants based on MicroCoder's technology stack, and improve the accuracy and stability of models in real business code scenarios through a precise evaluation framework and training experience.
• Educational programming aidsUsing MicroCoder's difficulty grading mechanism, we develop programming education tools for learners of different levels, providing progressive training support from basic to advanced levels.
• Code evaluation system upgradeExisting code evaluation platforms can integrate MicroCoder-Evaluator to solve the misjudgment problem caused by traditional strict matching, thereby improving evaluation accuracy and user experience.