Update README.md

README.md

@@ -1,91 +1,72 @@
+---
+pretty_name: YCbench
+dataset_name: ycbench
+language:
+  - en
+license: unknown          # Change to "mit", "cc-by-4.0", etc. if you pick one
+tags:
+  - finance
+  - startups
+  - ycombinator
+  - benchmarking
+  - tabular
+  - prediction
+  - venture-capital
+task_categories:
+  - tabular-classification
+  - text-classification
+size_categories:
+  - n<1K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: yc_w26_startups.csv
+      - split: train
+        path: yc_w26_traction.csv
+      - split: train
+        path: yc_w26_pre_demo_scores.csv
+      - split: train
+        path: yc_mentions.csv
+      - split: train
+        path: yc_mentions_early.csv
+---
+
 # YCbench
 
-**A dataset for benchmarking models that predict Y Combinator (YC) startup relative outperformance.**
+**A live benchmark dataset for forecasting startup outperformance in Y Combinator batches.**
 
-This dataset collects structured data related to Y Combinator startups, primarily focused on **W26** (Winter 2026) batch and related signals. It supports research and development of forecasting models that aim to identify which early-stage startups will **outperform their batch peers** in the short term (e.g., next 90 days), as explored in the [YC Bench](https://ycbench.com/) live benchmark.
+This dataset was introduced in the paper:  
+**["YC Bench: a Live Benchmark for Forecasting Startup Outperformance in Y Combinator Batches"](https://huggingface.co/papers/2604.02378)** by Mostapha Benhenda.
 
 ## Dataset Overview
 
-YCbench aggregates public and derived signals for YC companies, including:
-- Google/web mentions (as a proxy for visibility/traction)
-- Pre-demo day scores and predictions
-- Startup profiles
-- Traction metrics
-
-The data is useful for:
-- Training/evaluating ML models for startup success prediction
-- Cohort-relative benchmarking (comparing startups within the same YC batch)
-- Analyzing early signals of velocity and outperformance
-- Research on founder signals, traction, and short-term startup dynamics
+YCbench provides structured public signals for Y Combinator startups (focused on the **W26** batch of 196 companies). It enables rapid evaluation of models that predict which startups will **outperform their batch peers** in the short term (e.g., within ~90 days until Demo Day).
 
-## Files
+Instead of waiting years for exits or large funding rounds, performance is measured using a **Pre-Demo Day Score** that combines traction signals and web visibility.
 
-The dataset repository contains the following CSV files:
+### Key Files
 
-- `yc_mentions.csv` — Google/web mention counts for YC-related domains
-- `yc_mentions_early.csv` — Early-stage mention data
 - `yc_w26_startups.csv` — List of W26 startups with basic information
-- `yc_w26_traction.csv` — Traction metrics for W26 companies
+- `yc_w26_traction.csv` — Traction metrics
 - `yc_w26_pre_demo_scores.csv` — Pre-demo day scores, velocity scores, hybrid scores, etc.
+- `yc_mentions.csv` — Google/web mention counts
+- `yc_mentions_early.csv` — Early-stage mention data
 
-**Note:** These files have varying schemas (different columns). When loading with the `datasets` library, you may need to load them individually or as separate configurations rather than as a single default split.
-
-## Loading the Dataset
-
-You can load individual files using pandas or the Hugging Face `datasets` library with manual configuration:
+## Quick Load
 
 ```python
-import pandas as pd
 from datasets import load_dataset
+import pandas as pd
 
-# Load with pandas (recommended for now due to schema differences)
-df_mentions = pd.read_csv("hf://datasets/benstaf/ycbench/yc_mentions.csv")
-df_w26 = pd.read_csv("hf://datasets/benstaf/ycbench/yc_w26_startups.csv")
-
-# Or load specific files via datasets (if you define configs)
-# dataset = load_dataset("benstaf/ycbench", "w26_startups")  # example config
-```
-
-A preview of the data shows columns such as:
-- `domain` (string)
-- `google_mentions` (float)
-
-Other files include scores like `mentions_score`, `Velocity_Score`, `hybrid_score`, etc.
-
-## Motivation & Related Work
-
-This dataset is connected to **YC Bench** — a live benchmark that evaluates forecasting models on their ability to predict which YC startups will show the strongest execution velocity in the 90 days following the start of a batch.
-
-Key advantages of using YC batches for benchmarking:
-- Startups enter at roughly the same time and stage
-- Similar funding environment within a batch
-- Clean peer comparison (cohort-relative ranking)
-
-Instead of waiting years for exits or unicorn status, models are scored on measurable short-term outperformance.
-
-## Usage Ideas
-
-- Build baseline models using mention counts and traction signals
-- Experiment with hybrid scoring (mentions + velocity + other features)
-- Create leaderboards for LLM-based or traditional ML predictors
-- Analyze correlations between early signals and later performance
-
-## Limitations
-
-- Data is currently focused on W26 and general YC mentions
-- Column schemas are not fully unified across files
-- Some fields may contain null values
-- No official splits (train/validation/test) are provided yet
+# Load individual files (recommended for now)
+df_startups = pd.read_csv("hf://datasets/benstaf/ycbench/yc_w26_startups.csv")
+df_scores   = pd.read_csv("hf://datasets/benstaf/ycbench/yc_w26_pre_demo_scores.csv")
 
-## Contributing / Improvements
 
-Feel free to open a discussion or pull request if you want to:
-- Unify schemas into a single consistent dataset
-- Add more batches (S25, W27, etc.)
-- Include additional signals (LinkedIn activity, GitHub stars, funding news, etc.)
-- Add proper dataset cards, splits, or evaluation scripts
 
-## Related Links
+## Links
 
-- [YC Bench website](https://ycbench.com/) — Live leaderboard and benchmark
-- [YC Bench code on GitHub](https://github.com/benstaf/ycbench)
+- **Paper**: [arXiv 2604.02378](https://huggingface.co/papers/2604.02378)
+- **Live Benchmark**: [ycbench.com](https://ycbench.com/)
+- **GitHub**: [benstaf/ycbench](https://github.com/benstaf/ycbench)