DeepCoder-14B-Preview-RK3588-1.2.0
This version of DeepCoder-14B-Preview has been converted to run on the RK3588 NPU using w8a8 quantization.
This model has been optimized with the following LoRA:
Compatible with RKLLM version: 1.2.0
Useful links:
Pretty much anything by these folks:
marty1885 and
happyme531Original Model Card for base model, DeepCoder-14B-Preview, below:
DeepCoder-14B-Preview
🚀 Democratizing Reinforcement Learning for LLMs (RLLM) 🌟
DeepCoder Overview
DeepCoder-14B-Preview is a code reasoning LLM fine-tuned from DeepSeek-R1-Distilled-Qwen-14B using distributed reinforcement learning (RL) to scale up to long context lengths. The model achieves 60.6% Pass@1 accuracy on LiveCodeBench v5 (8/1/24-2/1/25), representing a 8% improvement over the base model (53%) and achieving similar performance to OpenAI's o3-mini with just 14B parameters.
Data
Our training dataset consists of approximately 24K unique problem-tests pairs compiled from:
- Taco-Verified
- PrimeIntellect SYNTHETIC-1
- LiveCodeBench v5 (5/1/23-7/31/24)
Training Recipe
Our training recipe relies on an improved version of GRPO (GRPO+) and iterative context lengthening, introduced in DeepScaleR.
GRPO+
We enhance the original GRPO algorithm with insights from DAPO to enable more stable training:
- Offline Difficulty Filtering: DAPO employs online dynamic sampling, discarding both entirely correct and entirely incorrect samples on the fly. While this helps maintain a more stable effective batch size, it introduces significant runtime overhead due to rejection sampling. Instead, we perform offline difficulty filtering on a subset of coding problems to ensure the training dataset remains within a suitable difficulty range.
- No Entropy Loss: We observed that including an entropy loss term often led to instability, with entropy growing exponentially and ultimately collapsing training. To mitigate this, we eliminate the entropy loss entirely.
- No KL Loss: Eliminating KL loss prevents the LLM from staying within trust region of the original SFT model. This removal also obviates the need to compute log probabilities for the reference policy, thereby accelerating training.
- Overlong Filtering (from DAPO): To preserve long-context reasoning, we mask the loss for truncated sequences. This technique enables DeepCoder to generalize to 64K-context inference despite being trained with a 32K context.
- Clip High (from DAPO): By increasing the upper bound in GRPO/PPO’s surrogate loss, we encourage more exploration and more stable entropy.
Iterative Context Lengthening
Our original Deepscaler-1.5B-Preview scaled long context training from 8K→16K→24K, achieving 33→38→43% on AIME respectively. Similarly, Deepcoder-14B-Preview is trained on 16K→32K, achieving 54→58% on LiveCodeBench (v5). DeepCoder-14B-Preview successfully generalizes to longer contexts when evaluated at 64K context, reaching 60.6%.
DeepCoder generalizes better to long contexts than the base distilled model, due to DAPO's overlong filtering. However, it's longer responses are often truncated when the max length is capped at 16K, which can lower its scores.
| Model | 16K | 32K | 64K |
|---|
| DeepCoder-14B-Preview | 45.6 | 57.9 | 60.6 |
| DeepSeek-R1-Distill-Qwen-14B | 50.2 | 53.0 | 53.0 |
A more detailed description of the training recipe can be found in our
blog post.
Evaluation
We evaluate Deepcoder-14B-Preview on various coding benchmarks, including LiveCodeBench (LCBv5), Codeforces, and HumanEval+.
| Model | LCB (v5)(8/1/24-2/1/25) | Codeforces Rating | Codeforces Percentile | HumanEval+ |
|---|
| DeepCoder-14B-Preview (ours) | 60.6 | 1936 | 95.3 | 92.6 |
| DeepSeek-R1-Distill-Qwen-14B | 53.0 | 1791 | 92.7 | 92.0 |
| O1-2024-12-17 (Low) | 59.5 | 1991 | 96.1 | 90.8 |
| O3-Mini-2025-1-31 (Low) | 60.9 | 1918 | 94.9 | 92.6 |
| O1-Preview | 42.7 | 1658 | 88.5 | 89 |
| Deepseek-R1 | 62.8 | 1948 | 95.4 | 92.6 |
| Llama-4-Behemoth | 49.4 | - | - | - |
Serving DeepCoder
Our model can be served using popular high-performance inference systems:
- vLLM
- Hugging Face Text Generation Inference (TGI)
- SGLang
- TensorRT-LLM
All these systems support the OpenAI Chat Completions API format.
Usage Recommendations
Our usage recommendations are similar to those of R1 and R1 Distill series:
- Avoid adding a system prompt; all instructions should be contained within the user prompt.
temperature = 0.6top_p = 0.95- This model performs best with
max_tokens set to at least 64000
License
This project is released under the MIT License, reflecting our commitment to open and accessible AI development.
We believe in democratizing AI technology by making our work freely available for anyone to use, modify, and build upon.
This permissive license ensures that researchers, developers, and enthusiasts worldwide can leverage and extend our work without restrictions, fostering innovation and collaboration in the AI community.
Acknowledgement
- Our training experiments are powered by our heavily modified fork of Verl, an open-source post-training library.
- Our model is trained on top of
DeepSeek-R1-Distill-Qwen-14B.
- Our work is done as part of Berkeley Sky Computing Lab and Berkeley AI Research.
Citation
1@misc{deepcoder2025,
2 title={DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level},
3 author={Michael Luo, Sijun Tan, Roy Huang, Ameen Patel, Alpay Ariyak, Qingyang Wu, Xiaoxiang Shi, Rachel Xin, Colin Cai, Maurice Weber, Ce Zhang, Li Erran Li, Raluca Ada Popa, Ion Stoica},
4 howpublished={\url{https://pretty-radio-b75.notion.site/DeepCoder-A-Fully-Open-Source-14B-Coder-at-O3-mini-Level-1cf81902c14680b3bee5eb349a512a51}},
5 note={Notion Blog},
6 year={2025}
7}
