Support multimodal input data

[moskomule]

Refer to #277.

Add parse_input_utterance and preprocessor to TemplateChatDataset to support multimodal input data.

• parse_input_utterance parses structured contents used in multimodal LMs
• preprocessor preprocesses each item, like image resizing

To set preprocessor from jsonnet, a base class Preprocessor is also prepared.

[Copilot]

Pull request overview

This PR adds support for multimodal input data (e.g., text + images) to flexeval, enabling evaluation of Vision Language Models (VLMs) and other multimodal language models. The implementation introduces two key features to TemplateChatDataset: parse_input_utterance to parse structured content from templates into lists of dictionaries (as required by OpenAI's multimodal API format), and preprocessor to preprocess items before template rendering (e.g., image resizing or base64 encoding).

Changes:

• Added parse_input_utterance parameter supporting literal_eval and json_loads parsing methods
• Added preprocessor parameter accepting a list of preprocessor instances for item transformation
• Created Preprocessor abstract base class defining the preprocessor interface
• Implemented ConvertImageToBase64 as an example preprocessor for image handling
• Added tests for the parse_input_utterance functionality

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 11 comments.

File	Description
flexeval/core/chat_dataset/template_based.py	Core implementation of multimodal support: added Preprocessor ABC, parse_input_utterance and preprocessor parameters to TemplateChatDataset and its subclasses
flexeval/multimodal/image_preprocessor.py	Example implementation of image-to-base64 preprocessor with resizing support
flexeval/multimodal/init.py	Module initialization exporting ConvertImageToBase64
tests/core/chat_dataset/test_template_based.py	Test coverage for parse_input_utterance feature with different parsing methods

---

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

[amanjainj98]

Adding the Preprocessor class to flexeval/core/chat_dataset/init.py would be helpful for usage.

flexeval/flexeval/core/chat_dataset/__init__.py

Line 5 in 4a8a31e

from .template_based import HFChatDataset, JsonlChatDataset, TemplateChatDataset, load_jinja2_template

[junya-takayama]

LGTM! nice feature 👏

Just a small nit. ↓