Allen Institute for AI researchers introduce SERA (Soft Verified Efficient repository Agents), a family of coding agents that is designed to be matched with much larger, closed systems by using only synthetic paths and supervised training.
What is SERA (Security and Resilience Act)?
SERA is AI2’s first Open Coding Agents release. SERA-32B – the flagship SERA model – is based upon Qwen 3’s 32B architecture. This is an agent that is designed to work at the repository level.
SERA 32B is able to resolve 49.5 percent of the SWE Bench Tested context at 32K. In 64K context, it is 54.2 percent. The numbers put it in the performance range of open weight systems like Devstral-Small-2 (with 24B parameters) and GLM 4.5 Air (with 110B parameters), while SERA stays fully open with code, weights, and data.
Hugging Face has released four SERA models, SERA-8B GA and SERA 32B GA.
Soft Verified Generation
Soft Verified Generation (SVG) is the basis of training. SVG generates agents trajectories which look like real developer workflows. Then, it uses the patch agreement between two rollsouts to softly signal correctness.
It is a:
- First Rollout: The sample function comes from a repository. GLM-4.6.1 in SERA 32B is a teacher model that receives bug styles or descriptions of changes. The GLM-4.6 can view and edit files and execute commands. The teacher model produces both a T1 trajectory and P1 patch.
- Synthetic Pull RequestThe system will convert the trajectory to a description similar to a real pull request. The text below summarizes the intent of each edit and how it will be implemented in a similar format to a real pull request.
- Second rolling outRather than seeing the entire pull request, the teacher only has the description. It creates new patch P2 and a trajectory T2. Both of these are aimed at implementing the change.
- Soft verificationLine by line, the patches P1 and are P2 will be compared. The recall score is calculated as the percentage of lines modified in P1 which appear in P2. The trajectory will be hard-verified when r = 1. The sample is only soft-verified for intermediate values.
Ablation studies have shown that strict validation is not necessary. Even when r is equal to 0, the performance of SWE Bench Verified remains similar for models trained using T2 trajectory with thresholds that are different. The results suggest that even noisy multistep traces are useful for coders.
The cost of training and data scale
SVG has been applied to 121 Python repositories that are derived from SWE-smith. SERA contains more than 200,000 paths from GLM-4 Air and GLM-4 teacher rolls.
SERA-32B was trained using a subset (25,00 T2 trajectory) of the Sera-4.6-Lite T2 data set. The training is done using Axolotl for standard fine tuning on Qwen-332B.
Many paths are longer than context limits. The team defines a truncation rate, which is the percentage of steps that can fit in 32K tokens. For the remainder, slices are selected with a high ratio of truncation. When comparing SWE Bench Verified scores, this ordered truncation outperforms the random truncation.
SERA 32B’s compute budget is reported at 40 GPU days, which includes both data generation and the training phase. The team calculated that using a scaling formula for performance and dataset size, SVG was 26 times less expensive than SkyRL Agent and SWE Smith to achieve similar SWE Bench scores.
Specialization in Repository
A key use case involves adapting an Agent to a particular repository. This is being studied by the research team on three SWE Bench Verified major projects: Django SymPy and Sphinx.
SVG produces 46,000-54,000 trajectory for each repository. Amounts of compute limit the number of trajectory training experiments per repository to 8,000. These trajectories are a mix between 3,000 soft verifiable T2 and 5,000 filtered filter T1 trajectories.
These specialized students are comparable to or slightly better than the GLM-4.5 Air teacher in 32K context. They also perform well compared with Devstral Small-2 for those repository subsets. In Django, the specialized student reaches a resolution rate of 52.23 % versus GLM 4.5-Air’s 51.20 %. SymPy’s specialized model achieves 51.11 percent, compared to 48.89 for GLM-4.5 Air.
The Key Takeaways
- SERA transforms coding agents to supervised learning problemsSERA-32B can be trained using standard fine-tuning on GLM-4.6 synthetic trajectory, without a reinforcement-learning loop or reliance on test suites from the repository.
- Soft Verified generation eliminates the requirement for testingSVG computes a soft-verification score using two rollsouts, patch overlap between P1 & P2, and researchers have shown that unverified or only weakly verifiable trajectories are still effective in training coding agents.
- Massive datasets of realistic agents from real repositoriesThis pipeline uses SVG on 121 Python Projects from the SWE Smith corpus. It produces more than 200,000 Trajectories, and creates one of the biggest open datasets available for coding agents.
- Effective training with cost and scale analysisSERA 32B train on T2 trajectories of 25,000 and according to the scale study, SVG costs about 26-times less than SkyRL Agent or 57-times less than SWE smith for similar SWE bench verified performance.
Click here to find out more Paper, Repo You can also find out more about the following: Model Weights. Also, feel free to follow us on Twitter Join our Facebook group! 100k+ ML SubReddit Subscribe Now our Newsletter. Wait! What? now you can join us on telegram as well.
