Test SERA in Claude Code for Free: https://github.com/allenai/sera-cli.

Technical Report: https://allenai.org/papers/opencodingagents

We support:

• Data generation from arbitrary personal codebases
• Data generation from existing code containers (SWE-Bench, SWE-smith, etc.)
• Data generation using open or closed source models
• SWE-agent and mini-swe-agent frameworks

Installation

Clone the repository locally, and then set up the environment.

With pip:

git clone --recurse-submodules https://github.com/allenai/SERA.git
cd SERA
conda create -n sera python=3.12
conda activate sera
pip install -e . -e modules/code2flow -e modules/SERA-SWE-Agent -e modules/SERA-mini-swe-agent

Existing Datasets

Best Subset (48000 Samples): https://huggingface.co/datasets/allenai/SERA-4.6-Lite-Best-Subset

• 48000 sample dataset of GLM 4.6 trajectories that result in SoTA 51.7% on SWE-Bench Verified at 32K context for 32B models.

GLM 4.6 Trajectories (72000 Samples):

• https://huggingface.co/datasets/allenai/Sera-4.6-Lite-T1 (36000)
• https://huggingface.co/datasets/allenai/Sera-4.6-Lite-T2 (36000)

GLM 4.5-Air Trajectories (139000 Samples):

• https://huggingface.co/datasets/allenai/Sera-4.5A-Full-T1 (72000)
• https://huggingface.co/datasets/allenai/Sera-4.5A-Full-T2 (66000)

GLM 4.5-Air Specialized Trajectories (132000 Samples from Django, Sympy, Sphinx):

• https://huggingface.co/collections/allenai/open-coding-agents-specialization

Training

See the README.md in sera/datagen/train.

Generation

The full configuration reference expands on the examples below.

Inference Servers

We support generation from open source and close source models.

We provide launch scripts for GLM-4.5-Air, GLM-4.6, and Qwen models.

Here is example usage:

# This sets TP to 8, launches a server on port 24444, and sets a seed of 42.
bash launch_glm45.sh 8 24444 42

The resulting server is usually http://HOSTNAME:PORT/v1 if its an openai server (which our servers are).

Single Server

Every experiment can be run either with one inference server, or multiple for higher efficiency. We release several examples showing how to reproduce our experiments or run your own. sera/config_schema.py contains a full list of configuration settings, enabling even more control over experiments.

1. Specialization to Django/Sympy from SWE-Bench

python sera/main.py \
    --config-name=specialization_django \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=URL

2. Specialization to Personal Repositories

sera/configs/specialization_personal.yaml defines a set of arbitrary codebases to generate data from. We use R2E-Gym as a toy repository to specialize to.

python sera/main.py \
    --config-name=specialization_personal \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=URL

Personal repositories require a little more involvement to generate data because we need to identify the main code folder, installation commands, etc. We suggest modifying the yaml file directly for this, instead of through CLI.

    - org_name: OpenHands
      last_name: OpenHands
      commits: # Provide exact commits to make containers on OR we automatically scrape some if not provided
        - 29b77be807e0e6aab380d953c0d79a166df4f0cc
        - cc8b677f3ec324fb7b9de86229f727b25741a66c
      install_cmds: # This is the default but sometimes personal repositories have their own installations
        - "pip install -e ."
      test_cmd: null
      python_version: 3.12 # For the docker container
      top_level_folder: # The main folder to look for to parse out functions.
        - openhands
    - org_name: R2E-Gym
      last_name: R2E-Gym
      install_cmds: 
        - "pip install -e ."

You can also set a Docker organization to push images to using generate.docker.docker_org=DOCKER_ORG. This makes it so created images are persistent. Make sure you have push permissions for the organization you choose. Otherwise, created images will be rebuilt every time the pipeline is rerun, taking a few extra minutes.

The default synthetic PRs created in the second rollout use SWE-Bench as demonstrations. In Personal PR Issues, we explain how you can set the demonstrations to be your own PR issues.

3. Using Closed-Source Models

If you want to use closed-source models, then the step of creating inference servers can be skipped.

python sera/main.py \
    --config-name=specialization_anthropic

4. Switching Frameworks

SWE-agent is the default harness but mini-swe-agent can be set as the harness. See the following toy example.

python sera/main.py \
    --config-name=specialization_django_anthropic \
    sweagent_cfgs=[mini_e2e,mini_e2e] \
    agent_harness=mini-swe-agent \
    distill.stage_one_config_name=mini_e2e \
    distill.stage_two_config_name=mini_e2e \
    name=test_mini_swe_agent

Multiple Servers

Multiple generation runs can be launched in parallel for the same experiment for large data generation runs using sharding. This is when the user defines multiple servers, and then shards the dataset to each server.

We use this for scaling swesmith to generate our largest datasets.

Replica 1:

python sera/main.py \
    --config-name=swesmith_scaling \
    distill.shard=0 \
    distill.total_shards=4 \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=URL_1

Replica 2:

python sera/main.py \
    --config-name=swesmith_scaling \
    distill.shard=1 \
    distill.total_shards=4 \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=URL_2

etc.

Creating Github Mirrors

This is a required step right now to create a docker container for your personal repository in the second example.

• Github Org Creation

We have created a mirror org on Github for everyone to use called oca-repos. However, any repositories mirrored here will be publicly viewable, so we recommend creating your own organization if you want to generate data from your private repository.

Continuing an Interrupted Run

At large scales, some generations (< 1%) will stall the teacher model, but this is enough to prevent the pipeline from completing a distillation step. To handle this, we support restarting any run from an arbitrary stage if the user chooses to kill a stalled run.

    stage_map = {
        "pipeline": -1,
        "generate": 0,
        "distill_stage_one": 1,
        "distill_stage_two": 2,
        "eval": 3,
        "postprocess": 4
    }

To continue a generation run simply make sure the name of the run (can be set via name=) matches the run to resume. Next, the argument stage=SOME_STAGE_MAP_KEY will continue the generation from whatever stage is chosen.

For example, if a few trajectories hang in distill_stage_one, you can run:

python sera/main.py \
    --config-name=specialization_django \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=URL \
    stage=distill_stage_two

And then the pipeline will skip the hanging trajectories and proceed to the second stage using only the successful rollouts from the first stage.

Alternatively, you can just rerun the initial command and as long as the experiment name matches the original run, it will pick up exactly where it left off instead of skipping to the next step in the pipeline.

Personal PR Issues

We write a script to scrape previous issue texts from any repository.

python scrape_github.py -o ORG_NAME -n REPO_NAME -c N_ISSUES

This saves a JSON file containing a list of issues to a pr_issues directory.

This list can then be passed into a run as:

distill.args.pipeline_repo=GENERATED_PATH

Full Configuration Reference

All settings are defined as dataclasses in sera/config_schema.py and can be overridden via the command line using OmegaConf dot notation. For example:

python sera/main.py --config-name=specialization_django \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=http://HOST:PORT/v1 \
    distill.sweagent_wrapper_config.num_workers=16 \
    eval.compare_patch_threshold=0.5

Top-Level Settings

Generate Settings (generate.*)

Personal Repos (generate.personal_repos)

Defined as a list in YAML config files. Each entry supports:

Existing Repos (generate.existing_repos)

For repositories with pre-built Docker containers (SWE-Bench, SWE-Smith, etc.). Defined as a list in YAML config files.

Distill Settings (distill.*)

Model (distill.model.*)

SWE-Agent Wrapper (distill.sweagent_wrapper_config.*)

Sharding & Stage Configs

Eval Settings (eval.*)

Postprocess Settings (postprocess.*)

Example: All Settings on CLI

python sera/main.py \
    --config-name=specialization_django \
    stage=pipeline \
    name=my_experiment \
    experiment_dir=./experiments \
    metadata_dir=./metadata \
    generate.fns_per_repo=5000 \
    generate.insts_per_fn=1 \
    generate.docker.docker_org=my-docker-org \
    generate.docker.gh_mirror_org=my-gh-mirrors \
    distill.model.name=openai/GLM-4.5-Air \
    distill.model.url=http://localhost:24444/v1 \
    distill.sweagent_wrapper_config.num_workers=24 \
    distill.sweagent_wrapper_config.per_instance_cost_limit=5.0 \
    distill.sweagent_wrapper_config.temperature=0.6 \
    distill.shard=0 \
    distill.total_shards=1 \
    eval.compare_patch_threshold=0.5 \
    postprocess.tool_call_format=hermes \
    postprocess.add_think=false \
    postprocess.enforce_submit=true

Citation

@misc{shen2026sera,
      title={SERA: Soft-Verified Efficient Repository Agents}, 
      author={Ethan Shen and Danny Tormoen and Saurabh Shah and Ali Farhadi and Tim Dettmers},
      year={2026},
      eprint={2601.20789},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.20789}, 
}