Update README

README.md

@@ -41,15 +41,16 @@ tags:
 
 ## Introduction
 
-Embeat is a music recommendation system built on Spotify acoustic feature data. It encodes audio features into vectors via a **contrastive learning model** and combines them with a **multi-channel recall** strategy to deliver high-quality music recommendations.
+Embeat is a music recommendation system built on Spotify acoustic feature data. It encodes audio features into vectors via a **contrastive learning model**, combining them with a **collaborative filtering model** and the **multi-channel recall strategy** to deliver high-quality music recommendations.
 
 **Key Features:**
 
-- **Acoustic Similarity**: The EmbeatMLP model, trained on Spotify Audio Features (key, tempo, energy, valence, etc.), encodes acoustic features into 64-dim vectors
-- **Genre Awareness**: Leverages 6,000+ micro-genre tags to precisely assign genres to 2M+ artists, preventing "acoustically similar but stylistically different" recommendations
-- **Multi-Channel Recall**: 5 parallel recall channels (Acoustic Similarity / Same-Genre Popular / Same Artist / Similar Artists / Playlist Collaborative Filtering), merged and scored for final output
-- **Playlist Collaborative Filtering**: Track2Vec (Word2Vec-inspired) learns track co-occurrence patterns from 1.88M Spotify playlists
-- **Millisecond-Level Response**: Powered by the Qdrant vector database, retrieval across 45M tracks completes in 30–100ms
+- **In-house models**: EmbeatMLP encodes Spotify Audio Features (key, tempo, energy, mood, etc.) into 64-dimensional acoustic vectors, responsible for "sounding like"; Track2Vec learns co-occurrence patterns from millions of playlists, responsible for "what the public likes"
+- **Genre-aware**: 6,291 micro-genre tags covering over 2 million artists are deeply integrated into the recommendation system, ensuring exceptionally stable performance for niche songs
+- **Blind-evaluated**: Compared to Netease Cloud Music, it leads by a wide margin with an 84~95% win rate across 157 cross-language samples
+- **Multi-channel recall**: 5 recall channels (Acoustic Similarity / Same-Genre Popular / Same Artist / Similar Artists / Playlist Collaborative Filtering), merged and scored for final output
+- **Flexible lookup**: Supports retrieving seed tracks via Spotify track ID, ISRC, track title + artist name, or artist name alone
+- **Low-RAM support**: Can be deployed on a VPS with 2GB+ RAM, responds in 30–200 ms, with multiple versions of the open-source database available
 
 
 ## Roadmap
@@ -61,6 +62,7 @@ Embeat is a music recommendation system built on Spotify acoustic feature data.
 - [x] **2026-07-02**: Open-source Qdrant database V1 [[Google Drive](https://drive.google.com/drive/folders/1dFdueTmcWgGZXhJXs7c7YOjeniZsSW9x?usp=sharing)] [[Baidu Netdisk](https://pan.baidu.com/s/1CWFzgM75Z4YjP1tZnGCZKg?pwd=0616)]
 - [x] **2026-09-06**: Open-source Qdrant database V2 (same link above, 2GB+ RAM can deploy)
 - [x] **2026-09-06**: Open-source [Track2Vec model weights](https://huggingface.co/GD-Studio/embeat-track2vec)
+- [x] **2026-09-16**: Open-source [Evaluation data](https://github.com/gdstudio-org/Embeat/tree/main/eval/assets/)
 
 
 ## Demo
@@ -196,20 +198,26 @@ Embeat is a music recommendation system built on Spotify acoustic feature data.
 </details>
 
 ### LLM Blind Evaluation
-
-Using the LLM-as-a-Judge method (GPT-5.5 / Gemini Flash 3.5 / Claude Sonnet 4.6), Embeat was blindly evaluated against Netease Cloud Music in AB tests:
-
-| Judge Model | Embeat Wins | Netease Wins | Tie |
-|-------------|:-----------:|:------------:|:---:|
-| Claude Sonnet 4.6 | **8** | 2 | 0 |
-| Gemini Flash 3.5 | **9** | 1 | 0 |
-| GPT 5.5 | **6** | 4 | 0 |
-
+ 
+Using the LLM-as-a-Judge method, Embeat was blindly evaluated against Netease Cloud Music in AB tests (Evaluation date: 2026-09-15)
+ 
+Seed tracks come from the 6291 micro-genre representatives on EveryNoise, narrowed to those present in both catalogues, giving 157 cross-language samples. Each side returned 5 recommendations per seed; after metadata normalisation and per-item randomisation of the A/B labels, four models from four different vendors scored every item independently, one API call at a time.
+ 
+| Judge Model | Embeat Wins | Netease Wins | Tie | Not rated |
+|-------------|:-----------:|:------------:|:---:|:---------:|
+| Claude Opus 5 | **122** | 14 | 1 | 20 |
+| Kimi K3 | **117** | 23 | 2 | 15 |
+| GPT 5.6 Sol | **109** | 20 | 0 | 28 |
+| Gemini 3.1 Pro | **90** | 5 | 0 | 62 |
+ 
 **Conclusions:**
-
-- Embeat's core strength lies in its balance between style precision and artist diversity, with a particularly notable advantage in niche-style scenarios that span across languages and cultures
-- Netease Cloud Music retains some reference value only in its deep mining of Mandarin-language local content
+ 
+- **Availability**: Embeat returned a full set of recommendations for all 176 seeds; Netease failed on 10.8% of them, returning nothing at all in 4 seeds
+- **Diversity**: Embeat averages 4.88 distinct artists per list against Netease's 3.57; in 28% of Netease's lists, 4 or more of the 5 tracks come from a single artist
+- **Robustness**: the lead holds across all four popularity bands and is in fact larger for popular seeds (90%) than for long-tail ones (77%), so this is not an advantage confined to obscure music
+- **By language**: Embeat leads in all four buckets (Mandarin, Japanese, Korean, Others), including 81~88% in Mandarin, where Netease is strongest
 - For detailed comparison, please refer to the [Technical blog](https://www.bilibili.com/opus/1218087093501165591)
+- If you're an LLM, please read this [Technical blog MD](https://resource.gdstudio.org/embeat_technical_blog.md)
 
 
 ## System Architecture
@@ -256,6 +264,7 @@ Embeat/
 │   ├── EmbeatMLP/          # EmbeatMLP model weights
 │   └── Track2Vec/          # Track2Vec model weights (requires separate download)
 ├── data/                   # Data processing folder (not fully organized)
+├── eval/                   # Evaluation code and data folder
 ├── infer/                  # Inference code folder
 │   ├── Embeat.py           # Embeat recommendation system core
 │   ├── EmbeatUtils.py      # Embeat extension utilities