[data] feat: Add dataset for Qwen-Image by chenyingshu · Pull Request #6 · zhtmike/verl

chenyingshu · 2026-01-09T03:04:29Z

What does this PR do?

Add dataset QwenDataset for Qwen-Image.
Add UT for dataset, testing dataloader and dataset configs.

Checklist Before Starting

Search for similar PRs. Paste at least one query link here: ...
Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
- {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
- If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
- {type} is in feat, fix, refactor, chore, test
- If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
- Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

Read the Contribute Guide.
Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
Add / Update the documentation.
Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

* add training engine * fix init * fix typs

* init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright

* init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn

zhtmike · 2026-01-09T03:07:53Z

verl/utils/dataset/qwen_dataset.py

+# 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.
+# ============================================================================


no need # ============================================================================

Add a comment on line R14`

Copilot

Pull request overview

This PR adds a new QwenDataset class for handling text prompts in Qwen-Image models, particularly for text-guided vision generation tasks. The dataset supports loading prompts from text files, tokenization with configurable templates, and extraction of ground truth data for OCR tasks.

Key changes:

New dataset implementation with prompt filtering, truncation, and tokenization support
Integration with existing reward loop for diffusion models
Unit tests for dataset functionality and dataloader compatibility

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
verl/utils/dataset/qwen_dataset.py	New dataset class for Qwen-Image with prompt loading, tokenization, and ground truth extraction
verl/utils/dataset/init.py	Export the new QwenDataset class
verl/experimental/reward_loop/reward_manager/diffusion.py	Remove unused commented code
tests/utils/dataset/test_qwen_dataset_on_cpu.py	Unit tests for QwenDataset with basic functionality and max_samples parameter
tests/experimental/reward_loop/test_diffusion_reward_model_genrm.py	Update test to use "input_ids" instead of "prompts" key

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

verl/utils/dataset/qwen_dataset.py

tests/utils/dataset/test_qwen_dataset_on_cpu.py

verl/utils/dataset/qwen_dataset.py

Copilot · 2026-01-09T03:10:49Z

verl/utils/dataset/qwen_dataset.py

+    Args:
+        data_files (str): Path to the text file containing prompts.
+        tokenizer (PreTrainedTokenizer): Tokenizer to tokenize the prompts.
+        config (OmegaConf): the data config.


The docstring says the parameter type is OmegaConf but the actual type hint is DictConfig. Consider updating the docstring to match the type hint for consistency.

Suggested change

config (OmegaConf): the data config.

config (DictConfig): the data config.

Copilot · 2026-01-09T03:10:49Z

verl/utils/dataset/qwen_dataset.py

+        if self.filter_overlong_prompts:
+            self.prompts = [x for x in self.prompts if len(x) <= self.max_prompt_length]
+
+        if self.max_samples > 0 and self.max_samples < len(self.prompts):
+            self.prompts = self.prompts[: self.max_samples]


The order of operations could be optimized. Currently, the code filters overlong prompts (line 62) before applying max_samples (lines 64-65). If filter_overlong_prompts significantly reduces the dataset size, the final dataset might have fewer samples than max_samples. Consider applying max_samples before filtering to ensure you get the requested number of samples, or document this behavior clearly.

tests/utils/dataset/test_qwen_dataset_on_cpu.py

verl/utils/dataset/qwen_dataset.py

zhtmike · 2026-01-09T03:13:36Z

if possible, we maintain the original configurations and returns from https://github.com/volcengine/verl/blob/main/verl/utils/dataset/rl_dataset.py#L70
and shows the minimal changes adapting to QwenDataset

chenyingshu · 2026-01-09T03:52:53Z

if possible, we maintain the original configurations and returns from https://github.com/volcengine/verl/blob/main/verl/utils/dataset/rl_dataset.py#L70 and shows the minimal changes adapting to QwenDataset

updated returns

zhtmike · 2026-01-09T08:10:18Z

verl/utils/dataset/qwen_dataset.py

+    def maybe_filter_out_long_prompts(self, prompts: list):
+        # filter out too long prompts
+        if self.filter_overlong_prompts:
+            prompts = [x for x in prompts if len(x) <= self.max_prompt_length]


I think the official filter_out_long_prompts are based on the length of the tokenized integers, not the length of the string.

* add entroypoint (#1) * add training engine (#2) * add training engine * fix init * fix typs * move folders & make for two-forward pass in training loop (#4) * Add diffusion reward loop (#3) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * [fix] update customized reward func in UT (#5) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn * init dataset for Qwen-Image * pass UT * update return, update UT * pass UT * align with rl_dataset * pass UT * update filter long prompts * debug * clean code --------- Co-authored-by: Cheung Ka Wai <zhtmike@gmail.com>

* add entroypoint (#1) * add training engine (#2) * add training engine * fix init * fix typs * move folders & make for two-forward pass in training loop (#4) * Add diffusion reward loop (#3) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * [fix] update customized reward func in UT (#5) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn * Update 20260109 (#8) * Update 20260109 * update * fix CI * [data] feat: Add dataset for Qwen-Image (#6) * add entroypoint (#1) * add training engine (#2) * add training engine * fix init * fix typs * move folders & make for two-forward pass in training loop (#4) * Add diffusion reward loop (#3) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * [fix] update customized reward func in UT (#5) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn * init dataset for Qwen-Image * pass UT * update return, update UT * pass UT * align with rl_dataset * pass UT * update filter long prompts * debug * clean code --------- Co-authored-by: Cheung Ka Wai <zhtmike@gmail.com> * add new config; debug actor * debug; add reward config; add adv, policy loss * debug reward loop * init diffusers engine UT * debug * debug * deubg actor forward * debug * merge * add UT for adv and loss * pass adv&loss UTs; pass engine backward UT * clean debug code --------- Co-authored-by: Cheung Ka Wai <zhtmike@gmail.com>

* add entroypoint (#1) * add training engine (#2) * add training engine * fix init * fix typs * move folders & make for two-forward pass in training loop (#4) * Add diffusion reward loop (#3) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * [fix] update customized reward func in UT (#5) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn * Update 20260109 (#8) * Update 20260109 * update * fix CI * [data] feat: Add dataset for Qwen-Image (#6) * add entroypoint (#1) * add training engine (#2) * add training engine * fix init * fix typs * move folders & make for two-forward pass in training loop (#4) * Add diffusion reward loop (#3) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * [fix] update customized reward func in UT (#5) * init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn * init dataset for Qwen-Image * pass UT * update return, update UT * pass UT * align with rl_dataset * pass UT * update filter long prompts * debug * clean code --------- Co-authored-by: Cheung Ka Wai <zhtmike@gmail.com> * update to align verl data format * debug --------- Co-authored-by: Cheung Ka Wai <zhtmike@gmail.com>

zhtmike and others added 7 commits January 6, 2026 16:04

add entroypoint (zhtmike#1)

de827c1

add training engine (zhtmike#2)

298c556

* add training engine * fix init * fix typs

move folders & make for two-forward pass in training loop (zhtmike#4)

d4218c7

Add diffusion reward loop (zhtmike#3)

03b166e

* init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright

[fix] update customized reward func in UT (zhtmike#5)

92d945b

* init reward; add ocr reward * update disrm input * add unit test * pass ut * fix typos/bugs * update copyright * update customized reward_fn

init dataset for Qwen-Image

f1060f7

pass UT

1a18e02

chenyingshu changed the title ~~Add dataset for Qwen-Image~~ [data] Add dataset for Qwen-Image Jan 9, 2026

zhtmike requested a review from Copilot January 9, 2026 03:06

Copilot started reviewing on behalf of zhtmike January 9, 2026 03:06 View session

chenyingshu changed the title ~~[data] Add dataset for Qwen-Image~~ [data] feat: Add dataset for Qwen-Image Jan 9, 2026

zhtmike reviewed Jan 9, 2026

View reviewed changes

Copilot AI reviewed Jan 9, 2026

View reviewed changes

chenyingshu added 2 commits January 9, 2026 11:48

update return, update UT

973ae13

pass UT

84a75b0

zhtmike force-pushed the verl-omni branch from 56b9dd7 to 0c75246 Compare January 9, 2026 06:38

chenyingshu added 2 commits January 9, 2026 15:03

Merge remote-tracking branch 'origin/verl-omni' into verl-omni-data

2ac9d8a

align with rl_dataset

8a98a08

zhtmike reviewed Jan 9, 2026

View reviewed changes

chenyingshu added 4 commits January 9, 2026 16:11

pass UT

cbb614e

update filter long prompts

0489547

debug

d8f8e95

clean code

44fa963

zhtmike merged commit ef70047 into zhtmike:verl-omni Jan 9, 2026
4 checks passed

chenyingshu deleted the verl-omni-data branch January 29, 2026 07:26

	config (OmegaConf): the data config.
	config (DictConfig): the data config.

Comments

Conversation

chenyingshu commented Jan 9, 2026

What does this PR do?

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

zhtmike Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI Jan 9, 2026

Choose a reason for hiding this comment

Uh oh!

chenyingshu Jan 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Jan 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

zhtmike commented Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chenyingshu commented Jan 9, 2026

Uh oh!

zhtmike Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

chenyingshu Jan 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

zhtmike Jan 9, 2026 •

edited

Loading

zhtmike commented Jan 9, 2026 •

edited

Loading

zhtmike Jan 9, 2026 •

edited

Loading