Be part of our every day and weekly newsletters for the newest updates and unique content material on industry-leading AI protection. Be taught Extra
Researchers at Collectively AI and Agentica have launched DeepCoder-14B, a brand new coding mannequin that delivers spectacular efficiency corresponding to main proprietary fashions like OpenAI’s o3-mini.
Constructed on high of DeepSeek-R1, this mannequin offers extra flexibility to combine high-performance code technology and reasoning capabilities into real-world purposes. Importantly, the groups have absolutely open-sourced the mannequin, its coaching information, code, logs and system optimizations, which can assist researchers enhance their work and speed up progress.
Aggressive coding capabilities in a smaller bundle
The analysis workforce’s experiments present that DeepCoder-14B performs strongly throughout a number of difficult coding benchmarks, together with LiveCodeBench (LCB), Codeforces and HumanEval+.
“Our mannequin demonstrates robust efficiency throughout all coding benchmarks… corresponding to the efficiency of o3-mini (low) and o1,” the researchers write in a weblog publish that describes the mannequin.
Apparently, regardless of being educated totally on coding duties, the mannequin exhibits improved mathematical reasoning, scoring 73.8% on the AIME 2024 benchmark, a 4.1% enchancment over its base mannequin (DeepSeek-R1-Distill-Qwen-14B). This means that the reasoning expertise developed by RL on code may be generalized successfully to different domains.
Probably the most placing side is attaining this stage of efficiency with solely 14 billion parameters. This makes DeepCoder considerably smaller and probably extra environment friendly to run than many frontier fashions.
Improvements driving DeepCoder’s efficiency
Whereas creating the mannequin, the researchers solved among the key challenges in coaching coding fashions utilizing reinforcement studying (RL).
The primary problem was curating the coaching information. Reinforcement studying requires dependable reward indicators indicating the mannequin’s output is appropriate. Because the researchers level out, “Not like math—the place plentiful high-quality, verifiable information is available on the Web—the coding area suffers from a relative shortage of such information.”
To deal with this drawback, the DeepCoder workforce carried out a strict pipeline that gathers examples from totally different datasets and filters them for validity, complexity and duplication. This course of yielded 24,000 high-quality issues, offering a strong basis for efficient RL coaching.
The workforce additionally designed an easy reward operate that solely supplies a optimistic sign if the generated code passes all sampled unit exams for the issue inside a selected time restrict. Mixed with the high-quality coaching examples, this outcome-focused reward system prevents the mannequin from studying methods like printing memorized solutions for public exams or optimizing for easy edge circumstances with out fixing the core drawback.
The mannequin’s core coaching algorithm relies on Group Relative Coverage Optimization (GRPO), a reinforcement studying algorithm that proved very profitable in DeepSeek-R1. Nonetheless, the workforce made a number of modifications to the algorithm to make it extra steady and permit the mannequin to proceed bettering because the coaching extends for an extended time.
Lastly, the workforce prolonged the mannequin’s context window iteratively, first coaching it on shorter reasoning sequences and regularly growing the size. Additionally they developed a filtering methodology to keep away from penalizing the mannequin when it created reasoning chains that exceeded the context limits when fixing a tough immediate.
The researchers clarify the core thought: “To protect long-context reasoning whereas enabling environment friendly coaching, we integrated overlong filtering… This method masks out truncated sequences throughout coaching in order that fashions aren’t penalized for producing considerate however prolonged outputs that exceed the present context restrict.”
The coaching was regularly scaled from a 16K to a 32K context window, and the ensuing mannequin might additionally clear up issues that required as much as 64K tokens.
Optimizing long-context RL coaching
Coaching massive fashions with RL, particularly on duties requiring lengthy generated sequences like coding or advanced reasoning, is computationally intensive and sluggish. A significant bottleneck is the “sampling” step, the place the mannequin generates probably 1000’s of tokens per instance within the batch. Variations in response size imply some responses end a lot later than others, leaving GPUs idle and slowing down the complete coaching loop.
To speed up this, the workforce developed verl-pipeline, an optimized extension of the open-source verl library for reinforcement studying from human suggestions (RLHF). The important thing innovation, which they name “One-Off Pipelining,” rearranges the response sampling and mannequin updates to cut back the bottlenecks and accelerator idle time.
Their experiments confirmed that one-off pipelining offered as much as a 2x speedup for coding RL duties in comparison with baseline implementations. This optimization was essential for coaching DeepCoder inside an inexpensive timeframe (2.5 weeks on 32 H100s) and is now open-sourced as a part of verl-pipeline for the neighborhood to make use of and construct upon.
Enterprise impression
The researchers have made all of the artifacts for coaching and working DeepCoder-14B obtainable on GitHub and Hugging Face below a permissive license.
“By absolutely sharing our dataset, code, and coaching recipe, we empower the neighborhood to breed our work and make RL coaching accessible to all,” the researchers write.
DeepCoder-14B powerfully illustrates a broader, accelerating development within the AI panorama: the rise of extremely succesful but environment friendly and brazenly accessible fashions.
For the enterprise world, this shift signifies extra choices and better accessibility of superior fashions. Slicing-edge efficiency is not solely the area of hyperscalers or these keen to pay premium API charges. Fashions like DeepCoder can empower organizations of all sizes to leverage refined code technology and reasoning, customise options to their particular wants, and securely deploy them inside their environments.
This development can decrease the barrier to entry for AI adoption and foster a extra aggressive and revolutionary ecosystem, the place progress is pushed by open supply collaboration.
