Metadata-Version: 2.4
Name: ui-rl
Version: 0.1.0
Summary: RLVR Finetuning for Computer Use models
Project-URL: Repository, https://github.com/TobiasNorlund/ui-rl
Project-URL: Issues, https://github.com/TobiasNorlund/ui-rl/issues
Author-email: Tobias Norlund <tobias@norlund.se>
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with the Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright 2026 Tobias Norlund
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is
           distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
License-File: LICENSE
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.12
Requires-Python: >=3.12
Requires-Dist: docker>=7.1.0
Requires-Dist: httpx>=0.28.1
Requires-Dist: kubernetes>=34.1.0
Requires-Dist: pillow>=11.3.0
Requires-Dist: torch<2.9
Requires-Dist: torchvision>=0.23.0
Requires-Dist: tqdm>=4.67.1
Requires-Dist: transformers>=4.57.3
Description-Content-Type: text/markdown

<div align="center">
    <img alt="ui-rl: Reinforcement Fine-tuning for Computer Use models" src="assets/uirl.svg">
</div>

---

[![Build](https://github.com/TobiasNorlund/ui-rl/actions/workflows/ci.yml/badge.svg)](https://github.com/TobiasNorlund/ui-rl/actions/workflows/ci.yml) [![License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)

`ui-rl` is a **framework for fine-tuning Computer Use agent models**.
It provides utilities for scalable training on verifiable UI tasks to improve model reliability on targeted domains and tasks.
This allows you to focus more on building verifiable tasks and optimizing model performance, and less on boilerplate code such as agent loops, rollout generations, torch data loading etc.

## The Why

Computer Use deployments require very robust model performance. You typically want (very close to) 100% success rate on your specific task(s), and never any fatal mistakes. Even the best models seldom offer that reliability for arbitrary tasks.
The best way to achieve such high reliability is to do task-specific fine-tuning. 

However, performing such fine-tuning is complex. A typical pipeline consists of starting from an open source model, performing some Supervised Fine-Tuning (SFT) until a decent success rate is achieved, and finally use Reinforcement Learning from Verified Reward (RLVR) for the last-mile performance boost.

Getting all this up-and-running requires significant setup, and is where `ui-rl` aims to simplify.


## Key features
 - Perform CUA rollouts at scale in custom task environments, using state-of-the-art open models
 - Rollout serialization/deserialization out of the box
 - CUA specific Data Augmentation techniques for SFT
 - Training framework agnostic — pick your favorite torch-compatible trainer
 - Pre-built torch `Dataset` implementations for rollouts — start training in a just few lines


## Installation

```bash
pip install ui-rl
```

## How to use:

We exemplify usage of `ui-rl` via common use cases:


### Use Case: Generate a rollout for a custom UI task

1. Start by building a containerized task environment. For an example, see [examples/simple_data_entry/env](examples/simple_data_entry/env) or [this blog post](https://medium.com/@tobias.norlund/does-computer-use-actually-make-sense-60f0449e3770) for a thorough walkthrough
2. Implement a `TaskSpec` that specifies the task prompt, and how to run it using a so called `SessionRuntime` (local docker container in this case).  

```python
from ui_rl.task import TaskSpec
from ui_rl.runtime.docker import DockerSessionRuntime

class SimpleDataEntryTaskSpec(TaskSpec[DockerSessionRuntime]):
    def __init__(self, rows: List[int]):
        self.rows = rows

    def get_task_instruction(self):
        return f"""Your task is to submit data from a spreadsheet (seen on the left) into a form (seen on the right). Specifically, the following rows (as numbered in the left margin) from the spreadsheet are to be submitted: {", ".join(str(i) for i in self.rows)}.
Note: You may need to scroll to make the row visible in the sheet.
The form has to be submitted separately for each row. When the form has been submitted, return to the form to submit the next row. 
Submit a row by selecting each cell individually, copy its content by sending keys "ctrl+c", select the target form text input and paste using "ctrl+v".
Finally, submit the form and continue with the next row. Only finish when all rows have been successfully submitted"""

    def create_session(self, runtime: DockerSessionRuntime) -> str:
        return runtime.create_session(
            image="ui-rl/simple-data-entry:latest",
        )

```

3. Start vLLM to use as inference engine, or use a cloud hosting. We will use the [UITARS 1.5 7B](https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B) model.

**Note:** You probably need a GPU with at least 40GB of VRAM for running this model

```bash
docker run -it --rm --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface -p 8000:8000 vllm/vllm-openai:latest "ByteDance-Seed/UI-TARS-1.5-7B" --limit-mm-per-prompt '{"image":10,"video":0}'
```

4. Instantiate the TaskSpec, and create a rollout object to manage and record the CUA model calls:

```python
import httpx
from ui_rl.models.uitars15.rollout import UITARS15_Rollout

# Create shared httpx client for communicating with vLLM and task environment
httpx_client = httpx.Client()

# Create a task spec for submitting row no 2 (the first data row) from the spreadsheet
task_spec = SimpleDataEntryTaskSpec(rows=[2])

# Create rollout object
rollout = UITARS15_Rollout(
    task_spec=task_spec,
    model_host="localhost:8000",
    model_name="ByteDance-Seed/UI-TARS-1.5-7B",
    httpx_client=httpx_client,
    max_images_in_context=10,
    max_tokens=200,
    temperature=0.1
)
```

5. Create a `SessionRuntime` that manages the rollout session containers. We'll run them locally using `DockerSessionRuntime` (but Kubernetes is supported via `KubernetesSessionRuntime`).

```python
runtime = DockerSessionRuntime(httpx_client=httpx_client)
```

6. Run and serialize the rollout

```python
from ui_rl.agent import run_cua_rollout

try:
    run_cua_rollout(
        task_spec=task_spec,
        rollout=rollout,
        runtime=runtime,
        max_steps=20,
    )
    print("Rollout finished successfully!")
    print("Progress:", rollout.progress)
    rollout.save("rollout_1.json")
except Exception as e:
    print(f"Error when generating rollout: {e}")
```

**Note:** A successful rollout has a "progress", a dict specifying task-specific progress data. For our example task, it contains:

```json
{
    "submitted_row_indices": [0],
    "num_incorrect_submissions": 0
}
```

This can later be used to compute a reward during RL training. 


### Use Case: Train on rollouts

Saved rollouts can be used directly for training. 
Perhaps the simplest case is when doing [Rejection Sampling](https://rlhfbook.com/c/10-rejection-sampling).
`ui-rl` comes with built in torch `Dataset`s for directly loading rollouts into trainable token sequences:

```python
from transformers import AutoProcessor
from ui_rl.models.uitars15.dataset import UITARS15_RolloutDataset

processor = AutoProcessor.from_pretrained("ByteDance-Seed/UI-TARS-1.5-7B")
ds = UITARS15_RolloutDataset(
    processor=processor,
    rollout_path="rollout_1.json"
)

print(ds[0])
```

Which will print something like:

```
{'input_ids': tensor([151644,   8948,    198,  ...,    272,    863, 151645]),
 'attention_mask': tensor([1, 1, 1,  ..., 1, 1, 1]),
 'labels': tensor([  -100,   -100,   -100,  ...,    272,    863, 151645]),
 'reward': tensor(0.),
 'pixel_values': tensor([[1.4340, 1.4340, 1.4340,  ..., 2.1459, 2.1459, 2.1459],
         [1.4340, 1.4340, 1.4340,  ..., 2.1459, 2.1459, 2.1459],
         [1.4340, 1.4340, 1.4340,  ..., 2.1459, 2.1459, 2.1459],
         ...,
         [1.8865, 1.8865, 1.8865,  ..., 2.1032, 2.1032, 2.1032],
         [1.7114, 1.7114, 1.7114,  ..., 2.0464, 2.0464, 2.1032],
         [1.8865, 1.8865, 1.9157,  ..., 2.1032, 2.1032, 2.1032]]),
 'image_grid_thw': tensor([[ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92],
         [ 1, 58, 92]])}
```

Here, there are some things worth noting:

 - The dataset iterates over _sequences_, all unique token sequences used by the agent throughout the rollout. This means that LLM completions that extends the same sequence is represented by a single sequence. If all LLM completions keep extending the same token sequence, then `len(ds)` will be 1.
 - The `labels` are constructed such that only generated tokens are trained.

Finally, a `reward_fn` can be provided in the constructor to support RL training, see code for more details:

```python
def reward_fn(rollout: UITARS15_RolloutDataset.Rollout) -> float:
    # Minor detail: The data starts on row 2 in the spreadsheet in this example
    if [row - 2 for row in rollout.task_spec["rows"]] == rollout.progress["submitted_row_indices"]:
        return 1.0
    else:
        return -1.0

ds = UITARS15_RolloutDataset(
    processor=processor,
    rollout_path="rollout_1.json",
    reward_fn=reward_fn
)
```

The reward is returned when iterating over the dataset.


### Use case: Generate multiple rollouts in parallel

`ui-rl` supports utilities for generating rollouts in parallel via `run_rollouts(...)`. It takes three arguments:

 - `strategy: ui_rl.RolloutStrategy`: Defines what task specs to run and in what order
 - `rollout_worker: ui_rl.RolloutWorker`: Defines how to run a rollout (e.g. creating a rollout object followed by `run_cua_agent(...)`, like above). 
 - `max_parallel: int`: How many rollouts to run in parallel.

Under the hood, `run_rollouts()` uses a `ProcessPoolExecutor` to run the rollout workers concurrently. We start by creating a `RolloutWorker` for our task and model: 

```python
import logging
from ui_rl import RolloutWorker, RolloutResult

class SimpleDataEntryRolloutWorker(RolloutWorker):
    _runtime = None

    @classmethod
    def init_worker(cls, log_queue: Queue):
        # init_worker is run at init for each process pool worker

        # Configures logging for multiprocessing, and creates a reusable 
        # httpx client for each process worker at cls._httpx_client
        super().init_worker(log_queue)

        # Initialize a docker runtime for this worker
        cls._runtime = DockerSessionRuntime(httpx_client=cls._httpx_client, session_timeout=60)

    def run(self, rollout_id: int, task_spec: SimpleDataEntryTaskSpec) -> RolloutResult:
        # `run` takes a rollout_id and task_spec, executes a rollout and returns a `RolloutResult` 
        # It is executed for each task_spec returned by the rollout strategy

        rollout = UITARS15_Rollout(
            task_spec=task_spec,
            model_host="localhost:8000",
            model_name="ByteDance-Seed/UI-TARS-1.5-7B",
            httpx_client=self._httpx_client,
            max_images_in_context=10,
            max_tokens=200,
            temperature=0.1
        )
        logging.info(f"Starting rollout for task: {task_spec}")
        try:
            run_cua_rollout(
                task_spec=task_spec,
                rollout=rollout,
                runtime=self._runtime,
                max_steps=20,
            )
            logging.info(f"Rollout {rollout_id} completed")

            # Save rollout
            result = RolloutResult(rollout_id, task_spec, rollout.progress, None)
            rollout.save(f"rollout_{rollout_id:03d}.json")
            return result

        except Exception as e:
            logging.error(f"Rollout {rollout_id} was stopped due to an error: {e}")
            return RolloutResult(rollout_id, task_spec, rollout.progress, e)
```

As we can see, the rollout worker simply wraps the rollout logic we created earlier. Now we are ready to run:

```python
from ui_rl import run_rollouts, FixedStrategy

# Rollout strategy for running our SimpleDataEntry task once for rows 2, 3 and 4
strategy = FixedStrategy(
    tasks=[SimpleDataEntryTaskSpec(rows=[i]) for i in range(2, 5)]
)

# Create rollout worker
worker = SimpleDataEntryRolloutWorker()

# Run
run_rollouts(
    strategy=strategy,
    rollout_worker=worker,
    max_parallel=3
)
```