Update README

README.md

@@ -1,217 +1 @@
 # DataSelectEnv — OpenEnv Environment for Data Curation in ML Training
-
-🚀 **Overview**
-
-DataSelectEnv is a reinforcement learning environment that simulates a real-world machine learning workflow:
-selecting high-quality training data under constraints.
-
-Agents must decide:
-- which data to select
-- how much to select
-- and which strategy to prioritize
-
-while balancing:
-- data quality
-- diversity
-- noise
-- and budget
-
----
-
-🎯 **Motivation**
-
-In real-world ML systems (e.g., at companies like Meta or Hugging Face), performance is not just about model architecture —
-it heavily depends on which data you train on.
-
-This environment models:
-- active learning
-- data filtering
-- noisy dataset handling
-- budget-constrained training
-
----
-
-🧠 **Core Idea**
-
-Instead of training a model once, the agent interacts step-by-step:
-1. Observe current model performance and dataset state
-2. Select a batch of data using different strategies
-3. Incrementally train the model (`partial_fit`)
-4. Receive reward based on improvement and data quality
-
----
-
-⚙️ **Environment Design**
-
-📦 **Dataset**
-- Generated using `sklearn.make_classification`
-- 1500 samples, 20 features
-- Structured clusters with controlled noise
-- Split:
-  - Seed: 100 samples (initial training)
-  - Validation: 200 samples
-  - Pool: remaining samples
-
-🔥 **Noise Simulation**
-- Some samples have corrupted labels
-- Noisy samples also have distorted features
-- They appear highly uncertain but harmful → creates a realistic trap
-
----
-
-🔁 **Interaction Loop**
-
-`reset()`
-- Initializes dataset and model
-- Trains model on seed data
-- Returns initial observation
-
-`step(action)`
-1. Normalize strategy weights
-2. Sample batch using:
-   - uncertainty
-   - diversity
-   - random
-3. Train model incrementally (`SGDClassifier.partial_fit`)
-4. Compute reward
-5. Update state
-
----
-
-📊 **Observation Space**
-
-```json
-{
-  "remaining_budget": int,
-  "diversity_score": float,
-  "noise_estimate": float,
-  "current_performance": float,
-  "samples_available": int
-}
-```
-
----
-
-🎮 **Action Space**
-
-```json
-{
-  "action_type": "select_batch | stop",
-  "batch_size": int,
-  "strategy_weights": {
-    "uncertainty": float,
-    "diversity": float,
-    "random": float
-  }
-}
-```
-
-- Weights are normalized internally
-- Enables continuous trade-offs, not discrete actions
-
----
-
-🏆 **Reward Function**
-
-The reward reflects learning progress + data quality trade-offs:
-- **Positive signal**:
-  - improvement in model performance
-  - diverse data selection
-- **Negative signal**:
-  - noisy data
-  - redundant samples
-  - excessive cost
-
----
-
-🧪 **Key Learning Dynamics**
-
-This environment models real ML behaviors:
-- 📉 **Diminishing returns** — repeated data gives less benefit
-- ⚠️ **Noise trap** — uncertain samples can be misleading
-- 🧩 **Diversity importance** — covering more data space improves learning
-- 💸 **Budget constraint** — forces efficient decisions
-
----
-
-🔥 **What Makes This Challenging**
-- No single strategy works
-- Uncertainty alone fails due to noise
-- Random is safe but suboptimal
-- Best performance requires balancing multiple signals
-
----
-
-📈 **Expected Behavior**
-
-| Strategy | Outcome |
-| :--- | :--- |
-| Balanced | Best performance |
-| Random | Moderate |
-| Uncertainty-only | Worst (fails on noisy data) |
-
----
-
-🛠️ **Tech Stack**
-- Python
-- NumPy
-- scikit-learn (`SGDClassifier`)
-- Pydantic (for typed models)
-
----
-
-📁 **Project Structure**
-
-```
-DataSelectEnv/
-├── env.py
-├── models.py
-├── sampling.py
-├── reward.py
-├── test_env.py
-```
-
----
-
-▶️ **How to Run**
-
-```bash
-python test_env.py
-```
-
----
-
-📌 **Current Status**
-- ✅ Core environment implemented
-- ✅ Stable training loop
-- ✅ Realistic reward dynamics
-- 🔄 Next: tasks + graders + OpenEnv API
-
----
-
-🧠 **Key Insight**
-
-> “We simulate data curation as a sequential decision-making problem where agents must balance uncertainty, diversity, and noise under budget constraints, using real incremental model training.”
-
----
-
-👥 **Team Notes**
-- Do NOT modify reward aggressively — current balance is tuned
-- Focus next on:
-  - tasks (easy/medium/hard)
-  - graders
-  - API + deployment
-
----
-
-📌 **Future Work**
-- OpenEnv API integration
-- Hugging Face Spaces deployment
-- Baseline agent evaluation
-- Advanced dataset scenarios
-
----
-
-🏁 **Goal**
-
-Build a realistic, non-trivial environment that can be used to evaluate intelligent data selection strategies in ML systems.