Update README.md

Updated project name and clarified project goals and tasks related to LLM-based data format conversion and metadata harmonization.

README.md

@@ -1,96 +1,86 @@
-# OpenMind Wrangler: AI Agents for Data Preparation in the AI 3.0 Era
+#  Open Data BIDSifier: LLM-based Data format conversion
 
-This is the repository for the exploratory project OpenMind Wranglers, inspired by the accelerating progress of AI 3.0 models and their potential to automate complex data engineering workflows for multi-study neuroimaging and clinical datasets.
+This is the repository for an exploratory project OpenMind Wranglers, inspired by the accelerating progress of LLMs and their potential to automate format conversion for multi-study neuroimaging and clinical datasets.
+
+## Motivation
 
 ## Project Aim
 
-The goal of this project is to explore whether AI agents (primarily LLM-based) can meaningfully assist in dataset wrangling tasks that precede model inference or training — tasks that are time-consuming yet essential to reproducible, large-scale neuroimaging research.
+The goal of this project is to explore if LLM-based workflow machines (sometimes calles "AI agents) can meaningfully assist in metadata harmonization, data format transformation and data preprocessing tasks that precede AI model inference or training. These tasks are time-consuming yet essential to reproducible, large-scale neuroimaging research.
 
-While the BIDS standard ensures interoperability on the metadata level, many preprocessing steps — such as volume normalization, quality control, outlier detection, or label encoding — remain manual or semi-automated. These steps form a bottleneck in data-driven science, especially when aggregating datasets across multiple studies.
+While the BIDS standard ensures interoperability, there are some datasets for which no BIDS annotation is available. This is a "dead data" which can not be used on-par with BIDS datasets.
 
-In this proof-of-concept, we aim to determine whether a coordinated system of AI agents can reliably execute these operations and produce AI-ready dataset collections, similar to those hosted on platforms like HuggingFace Datasets: OpenMind, with minimal human intervention.
+In this proof-of-concept, we aim to determine whether a coordinated system of AI agents can reliably execute these operations and produce AI-ready dataset collections, similar to those hosted on platforms like HuggingFace Datasets: [OpenMind](https://huggingface.co/datasets/AnonRes/OpenMind), with minimal human intervention.
 
 ## The Vision
 
-If successful, OpenMind Wrangler will serve as a foundation for a general-purpose AI data engineering assistant, capable of producing multi-study datasets that are immediately usable for machine learning and statistical analysis — without manual wrangling.
+If successful, Open Data Wrangler will serve as a foundation for harmonizing multi-study datasets from open data, making sure these are immediately usable for machine learning and statistical analysis.
 
 ## Rough Work Plan
 
-The project will proceed along four parallel tracks: agent design, dataset exploration, wrangling task automation, and evaluation.
-
-### 1. Agent Design
-
-We will prototype a multi-agent system where each agent handles a specific wrangling subtask (e.g., normalization, QC, encoding). Agents will communicate via a shared memory and task queue.
-
-Goals:
-
-Design a modular agentic framework (e.g., LangChain, CrewAI, or AutoGen)
-
-Define prompt templates and tools for each subtask
-
-Implement logging and reasoning traceability
-
-Resources:
-
-LangChain Agent Docs
-
-OpenAI Function Calling Guide
-
-### 2. Dataset Exploration
-
-OpenNeuro
-
-Tasks:
-
-Retrieve metadata and BIDS structures
-
-Generate schema summaries and compatibility maps
-
-Identify data types for downstream analysis
-
-### 3. Wrangling Task Automation
-
-The central track: enabling LLM-driven automation of the following operations:
+We will focus on data and metadata harmonization and evaluation of the results. For evaluation, a manual baseline would be prepared by humans.
 
-Volume normalization (using NiBabel / ANTsPy)
+### 1. Metadata harmonization
 
-Quality control report generation
+Zenodo and other data portals are rich on non-BIDS neuroimaging data. For a primer, these datasets are suggested: 
+- [UniToBrain](https://zenodo.org/records/5109415)
+- [Cranial CT of 1 patient](https://zenodo.org/records/16816)
+- [BraTS 2020](https://www.kaggle.com/datasets/awsaf49/brats2020-training-data)
+- [Macaque neurodevelopment database](https://data.kitware.com/#collection/54b582c38d777f4362aa9cb3)
 
-Outlier detection (statistical and visual)
+The goal is to harmonize them with the following BIDS datasets:
+- [Sexing the parental brain in shopping: an fMRI study](doi:10.18112/openneuro.ds006844.v1.0.0)
+- [Cross-modal Hierarchical Control](https://openneuro.org/datasets/ds006628/versions/1.0.1)
 
-Label harmonization and encoding
+and really any other OpenNeuro dataset.
 
-Data documentation generation (Markdown / JSON-LD)
+6 WPs are the following
+Harmonizing **metadata** *with LLM based tools*:
+1. Annotation column names (from non-BIDS to BIDS)
+2. File structure
+3. Study metadata (fetching from repository HTMLs too)
 
-The focus is not on achieving perfection, but on evaluating how well an AI agent can assist or autonomously perform these tasks, given contextual metadata and goals.
+For the LLM-assisted workflow, following **tools** are suggested: Github Copilot in VS Code, LLMAnything. 
+Pick **any LLM** really, smaller LLMs tend to hallucinate more, therefore it is more interesting if they can make it too!
+Suggestions bigger LLMs: GPT-5, Claude, Kimi-K2, DeepSeek-R1
+Suggestions smaller LLMs: SmoLM, LLaMA-7B, Qwen-7B
 
-### 4. Evaluation and Benchmarking
+Harmonizing **metadata** *by hand*:
+4. Annotation column names (from non-BIDS to BIDS)
+5. File structure
+6. Study metadata (fetching from repository HTMLs too)
 
-The evaluation phase will assess the efficiency, accuracy, and robustness of agentic wrangling workflows compared to traditional, human-coded pipelines.
+For the manual harmonization, an IDE with Python / R is useful; as well as OpenRefine, an open-source tool for working with tabular data.
 
-Metrics will include:
 
-Time saved per dataset compared to manual pipelines
+Record the problems and the working time for both manual and LLM assisted harmonization.
 
-Accuracy of normalization / encoding
+Time planned: ~5 hours working time are planned for this step.
 
-Error detection rate (false positives / negatives in QC)
+### 2. Evaluation of the harmonized metadata
 
-Consistency across heterogeneous datasets
+Let's see how well the harmonization went!
+Use [BIDS validator](https://bids-standard.github.io/bids-validator/) on the newly-BIDS datasets, and report whether:
+1. The manual harmonization is BIDS compliant.
+2. The semi-automatic harmonization is BIDS compliant.
 
-We will also benchmark against a simple scripted baseline (e.g., manual nipype or pandas pipeline).
+3. Assess the differences in the harmonized metadata.
+4. Try to "stack" BIDS converted datasets on old BIDS datasets, and report the errors.
 
-Milestones
+Time planned: ~4 hours.
 
-✅ Literature & tooling review on AI-assisted data wrangling
+### 3. Try in action
 
-⚙️ Prototype LLM agent framework for dataset preprocessing
+Use [ResEncL](https://huggingface.co/AnonRes/ResEncL-OpenMind-MAE) or another model of your choice with JAX, and feed it the resulting harmonized dataset. Analyse and record the bugs. 
 
-🧠 Evaluation on 2–3 public BIDS datasets
+Time planned: ~4 hours.
 
-📊 Quantitative + qualitative benchmarking report
+### Teams and Participants / Skills
 
+Anyone from master students to experienced scientists is welcome to join. The project will involve a lot of tabular data analysis and scripting, for this, coding experience in Python and R and Shell experience are useful.
+For working with LLM agents, prompt engineering skills can be useful, though can also be acquired in this project.
 
-Generative augmentation: Can the agent propose synthetic data to fill gaps or balance classes?
+### References / Prior reading
 
-Teams and Participants
+- [What is an AI agent](https://blog.langchain.com/how-to-think-about-agent-frameworks/)
+- [Prompting guide from Meta](https://www.llama.com/docs/how-to-guides/prompting/)