DeepCoder 1.5B Preview

Overview

DeepCoder-1.5B-Preview is an open-source generative AI model tailored for code reasoning tasks and optimized for long-context understanding. Developed by Agentica, this model extends the capabilities of reinforcement learning for large language models (RLLMs), building upon DeepSeek-R1-Distilled-Qwen-1.5B as its foundational architecture. Through its innovative training methodology and rigorous dataset curation, DeepCoder-1.5B-Preview aims to advance the state-of-the-art in verifiable program synthesis and automated code analysis. For further background, details on this model are provided via the official model documentation and an in-depth technical blog post.

Model Architecture and Training Methodology

DeepCoder-1.5B-Preview is fine-tuned from the DeepSeek-R1-Distilled-Qwen-1.5B model, employing an advanced variant of the GRPO (Generalized Reward Policy Optimization) algorithm known as GRPO+. This methodology stems from the reinforcement learning paradigm and incorporates iterative context lengthening as introduced in DeepScaleR. The GRPO+ strategy improves stability and efficiency by integrating offline difficulty filtering, dispensing with entropy and KL losses, and leveraging techniques such as overlong filtering and surrogate loss clipping, as detailed in the DAPO research.

Iterative context lengthening allows the model to learn from shorter input sequences before gradually extending to longer ones. In practice, DeepCoder-1.5B-Preview was trained across sequential context lengths from 8K to 16K and finally to 24K tokens. This enables the model to generalize reasoning capabilities over contexts as long as 64K tokens during inference.

Data Sources, Filtering, and Verification

To ensure high-quality training and verifiability, DeepCoder-1.5B-Preview draws from approximately 24,000 unique problem-test pairs, integrating datasets such as Taco-Verified, PrimeIntellect SYNTHETIC-1, and LiveCodeBench v5. A multi-stage filtering pipeline is employed. First, problems are programmatically verified against their official solutions, retaining only cases where all standard unit tests (minimum of five per problem) pass, thereby minimizing the possibility of reward exploitation or ambiguous evaluation. Deduplication further ensures that each problem in the training set is unique.

The reward computation process operates within isolated code sandboxes for security and determinism. The Together Code Interpreter facilitates efficient, parallelized code evaluation, while a local subprocess sandbox aligns with LiveCodeBench's standard evaluation framework. The reward function is implemented as a sparse Outcome Reward Model: a generated solution receives a reward of 1 only if it passes all of up to 15 challenging unit tests; otherwise, the reward is 0. Each test is subject to a strict timeout, usually between 6 and 12 seconds.

System Optimization and Computational Pipeline

To accelerate RL-based training, DeepCoder leverages a modified 'verl-pipeline' system—an optimized extension to the Verl post-training library. These optimizations include one-off pipelining, which separates sampling and training into successive alternating steps to maximize throughput and minimize idle time during trainer and reward calculation.

This approach enables near-complete masking of reward computation and trainer idle times, reducing overall training durations by a reported factor of 1.4x for mathematical reasoning and up to 2x for code generation workloads.

Performance Metrics and Benchmarking

DeepCoder-1.5B-Preview has been benchmarked on several key evaluation suites, including LiveCodeBench (LCBv5), Codeforces, and HumanEval+. On LCBv5, the model achieves a score of 25.1, with a Codeforces rating of 963, placing it in the 28.5 percentile, and a HumanEval+ success rate of 73.0%. By comparison, its base model (DeepSeek-R1-Distilled-Qwen-1.5B) attains an LCBv5 score of 16.9 and a HumanEval+ score of 58.3%, underscoring a marked improvement attributable to the reinforcement learning and context lengthening strategies. Detailed evaluation logs and further breakdowns are available from model resources.

Model Family and Comparative Analysis

DeepCoder-1.5B-Preview is part of a broader family. A closely related model, DeepCoder-14B-Preview, is fine-tuned from Deepseek-R1-Distilled-Qwen-14B and achieves a LiveCodeBench Pass@1 accuracy of 60.6% at 64K context—a performance level comparable to other frontier reasoning models at similar or greater scale. Both models employ the same principles of iterative context length expansion and GRPO+ reinforcement learning, but differ in model size and eventual context generalization capacity. Additional details, including context scaling and comparisons with peer models, can be found in the comprehensive technical blog.

Limitations and Licensing

While DeepCoder-1.5B-Preview demonstrates robust long-context reasoning, there are noted trade-offs at shorter context lengths. For instance, average response lengths can exceed standard truncation thresholds at 16K tokens, leading to the potential for response truncation and corresponding reductions in benchmark scores. These effects are discussed further in model documentation.

DeepCoder-1.5B-Preview is released under the MIT License, with the aim of promoting open access and scientific transparency in AI research.

External Resources

• Agentica DeepCoder Model Card on Hugging Face
• DeepCoder Notion Technical Blog
• LiveCodeBench Benchmark Site
• Together Code Interpreter Documentation
• Verl-Pipeline for RL Pipelines
• Reinforcement Learning for LLMs (RLLM) GitHub
• DeepCoder-14B-Preview on Hugging Face
• DeepCoder-Preview-Dataset on Hugging Face
• Evaluation Logs and Reports (Google Drive)