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

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

Setting	Type	Default	Description
stage	str	"pipeline"	Which stage to run. Options: pipeline (all stages), generate, distill_stage_one, distill_stage_two, eval, postprocess
name	str | None	None	Name of the run. Creates a folder at experiment_dir/name. Auto-generated if not set. Used to resume interrupted runs.
experiment_dir	str	"./experiments"	Where to save experiment data (trajectories, rollouts, outputs)
metadata_dir	str	"./metadata"	Where to save parsed codebase graphs and other metadata
sweagent_cfg_dir	str	"./sera/configs/sweagent/"	Directory containing full SWE-agent config YAML files
sweagent_cfgs	list[str]	["e2e", "mini_e2e"]	Which SWE-agent configs to load into the experiment

Generate Settings (generate.*)

Setting	Type	Default	Description
generate.fns_per_repo	int	5000	Max number of functions to extract per repository
generate.insts_per_fn	int	1	Number of times to process each function through the pipeline. Increase to generate more samples.
generate.repo_parent_dir	str	"./repos"	Where to store cloned repositories
generate.docker.docker_org	str | None	None	Docker organization to push created images to. Makes containers persistent across reruns.
generate.docker.gh_mirror_org	str | None	None	GitHub mirror organization for personal repos (required for personal repo containerization)

Personal Repos (generate.personal_repos)

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

Setting	Type	Default	Description
org_name	str	required	GitHub organization name
last_name	str	required	GitHub repository name
commits	list[str] | None	None	Exact commits to create containers for. If not set, auto-scrapes n_commits commits.
n_commits	int	5	Number of commits to auto-scrape if commits is not specified
lookback	int	365	How many days to look back when auto-scraping commits
language	str	"python"	Repository language. Only Python is supported currently.
install_cmds	list[str]	["python -m pip install -e ."]	Commands to install the repository inside the container
test_cmd	str | None	None	Optional test command to verify installation
python_version	str	"3.10"	Python version for the Docker container
skip_package_name	list[str]	[]	Packages to skip installing (sidesteps rare dependency errors)
top_level_folder	list[str]	[]	Main code folder(s) to parse (e.g. src). Auto-detected if empty.
overwrite_cg	bool	False	Set True to regenerate codebase graphs instead of using cache
max_folder_depth	int	3	How deep to parse into the codebase. Higher = more functions extracted.

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.

Setting	Type	Default	Description
org_name	str	required	GitHub organization name
last_name	str	required	GitHub repository name
base_commit	str | None	None	Commit the container is based on. Auto-set for swebench/swesmith sources.
instance_id	str | None	None	SWE-Bench instance ID (e.g. django__django-7530). Used to auto-set base_commit.
source	str | None	None	Container source: "swebench", "swesmith", or leave empty for custom
image_name	str | None	None	Custom Docker image name (for repos that are not from SWE-Bench or SWE-Smith)
top_level_folder	list[str]	[]	Main code folder(s) to parse. Auto-detected if empty.
overwrite_cg	bool	False	Set True to regenerate codebase graphs instead of using cache
max_folder_depth	int	3	How deep to parse into the codebase. Higher = more functions extracted.

Distill Settings (distill.*)

Model (distill.model.*)

Setting	Type	Default	Description
distill.model.name	str	""	Model identifier. Use openai/ prefix for OpenAI-compatible servers (e.g. openai/GLM-4.5-Air), anthropic/ prefix for Anthropic API (e.g. anthropic/claude-sonnet-4-20250514).
distill.model.url	str | None	""	API endpoint URL (e.g. http://HOST:PORT/v1). Leave empty/null for official OpenAI or Anthropic APIs.

SWE-Agent Wrapper (distill.sweagent_wrapper_config.*)

Setting	Type	Default	Description
distill.sweagent_wrapper_config.num_workers	int	32	Number of concurrent rollouts
distill.sweagent_wrapper_config.per_instance_call_limit	int	115	Max number of rollout steps per instance
distill.sweagent_wrapper_config.per_instance_cost_limit	float	0.0	Max cost per rollout. Set to 0.0 for local models, > 0.0 for API models.
distill.sweagent_wrapper_config.total_cost_limit	float	0.0	Max total cost across all rollouts
distill.sweagent_wrapper_config.temperature	float	0.6	Model sampling temperature

Sharding & Stage Configs

Setting	Type	Default	Description
distill.shard	int	0	Current shard index (0-indexed) for parallel multi-server runs
distill.total_shards	int	1	Total number of shards to split the data into
distill.stage_one_config_name	str	"e2e"	SWE-agent config name for stage 1 rollouts. Must be in sweagent_cfgs.
distill.stage_two_config_name	str	"e2e"	SWE-agent config name for stage 2 rollouts. Must be in sweagent_cfgs.
distill.args	dict	{"pipeline": True, "pipeline_yaml": "sera/configs/pipeline/default_pipeline.yaml"}	Extra args passed to SWE-agent. Use distill.args.pipeline_repo=PATH to provide custom PR issues.

Eval Settings (eval.*)

Setting	Type	Default	Description
eval.compare_patch_threshold	float	1	Verification threshold. 1 = hard verification (exact patch match), 0 < r < 1 = soft verification, 0 = no verification.

Postprocess Settings (postprocess.*)

Setting	Type	Default	Description
postprocess.tool_call_format	str	"hermes"	Tool call format in output data. Options: hermes, xml, raw
postprocess.add_think	bool	False	Add <think> tags to output. Useful for training Qwen3 models when the teacher doesn't produce think tokens (e.g. Claude).
postprocess.add_train_key	bool	True	Add train key to assistant messages for axolotl, ensuring only assistant messages are trained on
postprocess.include_tool_json	bool	True	Include OpenAI-formatted tool JSON as a field in each sample (helps debugging)
postprocess.reformat_assistant_message	str | None	"keep_only_think"	How to handle multi-part assistant messages (e.g. <think>TEXT</think>MORE TEXT). Options: "" (keep all), "keep_only_think", "keep_only_non_think"
postprocess.enforce_submit	bool	True	Only process trajectories that successfully submitted, filtering out ones that hit cost/context limits

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

Training

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

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}, 
}