Upload 2 files

README.md

@@ -3,15 +3,14 @@
   <p align="center">
     <strong>An Open Foundation Model and Benchmark to Accelerate Generative Recommendation</strong>
   </p>
-
   <p align="center">
-    <a href="https://huggingface.co/Onerec">
+    <a href="https://huggingface.co/OpenOneRec">
         <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-OneRec-ffc107?color=ffc107&logoColor=white" />
     </a>
-    <a href="https://github.com/OnerecLM/OpenOneRec">
+    <a href="https://github.com/Kuaishou-OneRec/OpenOneRec">
         <img alt="GitHub Code" src="https://img.shields.io/badge/GitHub-OpenOneRec-black?logo=github" />
     </a>
-     <a href="">
+     <a href="https://arxiv.org/abs/2508.20900">
         <img alt="Paper" src="https://img.shields.io/badge/Paper-ArXiv-b31b1b?logo=arxiv" />
     </a>
     <a href="#license">
@@ -19,7 +18,6 @@
     </a>
   </p>
 </div>
-
 <br>
 
 ## 📖 Introduction
@@ -27,26 +25,26 @@
 **OpenOneRec** is an open-source framework designed to bridge the gap between traditional recommendation systems and Large Language Models (LLMs). While Generative Recommendation has shown promise, existing models often struggle with isolated data silos and a lack of reasoning capabilities.
 
 To address this, we introduce a unified framework that comprises:
-* **RecIF-Bench**: The first holistic Recommendation Instruction-Following Benchmark, containing **120M interactions** from 200k users across heterogeneous domains (Short Video, Ad, Product).
-* **OpenOneRec-Foundation Models**: A family of models (1.7B & 8B) built on the Qwen backbone. These models are trained on hundreds of billions of tokens, integrating collaborative signals with general semantics.
+* **RecIF-Bench**: The first holistic Recommendation Instruction-Following Benchmark, containing **100M interactions** from 200k users across heterogeneous domains (Short Video, Ads, Product).
+* **OneRec-Foundation Models**: A family of models (1.7B & 8B) built on the Qwen backbone. These models are trained on hundreds of billions of tokens, integrating collaborative signals with general semantics.
 * **Full-Stack Pipeline**: We open-source our comprehensive training pipeline, including data processing, co-pretraining, and post-training, to ensure full reproducibility and facilitate scaling law research in recommendation.
 
 ## 🔥 News
 
-* **[2025.xx.xx]** 🎉 **OpenOneRec-Foundation** models (1.7B, 8B) are now available on Hugging Face!
-* **[2025.xx.xx]** 📑 The technical report [OpenOneRec Technical Report] has been released.
-* **[2025.xx.xx]** 🚀 **RecIF-Bench** dataset and evaluation scripts are open-sourced.
+* **[2026.1.1]** 🎉 **OneRec-Foundation** models (1.7B, 8B) are now available on Hugging Face!
+* **[2026.1.1]** 📑 The technical report has been released.
+* **[2026.1.1]** 🚀 **RecIF-Bench** dataset and evaluation scripts are open-sourced.
 
 ## 📊 RecIF-Bench
 
-We propose **RecIF-Bench**, a comprehensive benchmark designed to rigorously evaluate recommendation foundation models. It organizes 8 distinct tasks into a four-layer capability hierarchy:
+We propose **RecIF-Bench** to rigorously assess the synergy between instruction following and domain-specific recommendation. It organizes 8 distinct tasks into a four-layer capability hierarchy:
 
 * **Layer 0: Semantic Alignment** (Item Understanding) 
-* **Layer 1: Fundamental Recommendation** (Short Video Rec, Ad Rec, Product Rec, Label Prediction) 
+* **Layer 1: Fundamental Prediction** (Short Video Rec, Ad Rec, Product Rec, Label Prediction) 
 * **Layer 2: Instruction Following** (Interactive Rec, Label-Conditional Rec) 
 * **Layer 3: Reasoning** (Recommendation Explanation) 
 
-The benchmark aggregates data from three domains: **Short Video** (Content), **Ad** (Commercial), and **Product** (E-commerce).
+The benchmark aggregates data from three domains: **Short Video** (Content), **Ads** (Commercial), and **Product** (E-commerce).
 
 ## 🤖 Model Zoo
 
@@ -54,10 +52,10 @@ The OpenOneRec-Foundation series is built upon the Qwen architecture, enhanced w
 
 | Model | Backbone | Parameters | Description | Link |
 | :--- | :--- | :--- | :--- | :--- |
-| **OneRec-Open-1.7B** | Qwen3-1.7B | 1.7B | Standard version trained on open-source data (~100B tokens) | [HuggingFace](https://huggingface.co/Onerec) |
-| **OneRec-Open-8B** | Qwen3-8B | 8B | Standard version trained on open-source data (~100B tokens) | [HuggingFace](https://huggingface.co/Onerec) |
-| **OneRec-Pro-1.7B** | Qwen3-1.7B | 1.7B | Enhanced version with proprietary tokens | *Coming Soon* |
-| **OneRec-Pro-8B** | Qwen3-8B | 8B | Enhanced version with proprietary tokens | *Coming Soon* |
+| **OneRec-Open-1.7B** | Qwen3-1.7B | 1.7B | Standard version trained on open-source data (~33B tokens) | [HuggingFace](https://huggingface.co/OpenOneRec/OneRec-1.7B) |
+| **OneRec-Open-8B** | Qwen3-8B | 8B | Standard version trained on open-source data (~33B tokens) | [HuggingFace](https://huggingface.co/OpenOneRec/OneRec-8B) |
+| **OneRec-Pro-1.7B** | Qwen3-1.7B | 1.7B | Scaled-up version with expanded datasets (~130B tokens) | [HuggingFace](https://huggingface.co/OpenOneRec/OneRec-1.7B-pro) |
+| **OneRec-Pro-8B** | Qwen3-8B | 8B | Scaled-up version with expanded datasets (~130B tokens) | [HuggingFace](https://huggingface.co/OpenOneRec/OneRec-8B-pro) |
 
 ## 🏗️ Method & Architecture
 
@@ -68,14 +66,14 @@ To bridge the modality gap, we treat items as a distinct modality using **Itemic
 
 ### 2. Training Pipeline
 Our framework utilizes the following recipe:
-* **Pre-Training**: Integrates collaborative signals via Itemic-Text Alignment and mixed-domain Co-Pretraining.
+* **Pre-Training**: Integrates collaborative signals via Itemic-Text Alignment and Full-Parameter Co-Pretraining.
 * **Post-Training**:
-    * *Stage 1*: Cold-Start SFT for basic instruction following.
-    * *Stage 2*: Alternating On-Policy Distillation & SFT to balance general reasoning and recommendation performance.
-    * *Stage 3*: Recommendation-oriented Reinforcement Learning (Rec-RL).
+    * *Stage 1*: Multi-task Supervised Fine-tuning for basic instruction following.
+    * *Stage 2*: On-policy Distillation to restore general reasoning performance.
+    * *Stage 3*: Reinforcement Learning to enhance recommendation capabilities.
 
 <div align="center">
-  <img src="assets/framework.png" width="90%" alt="OpenOneRec Overall Framework" />
+  <img src="assets/framework.png" width="80%" alt="OpenOneRec Overall Framework" />
   <br>
   <em>Figure: The Overall Framework of OpenOneRec.</em>
 </div>
@@ -85,21 +83,35 @@ Our framework utilizes the following recipe:
 ### Results on RecIF-Bench
 OpenOneRec-Foundation achieves **State-of-the-Art (SOTA)** results across RecIF-Bench tasks, significantly outperforming baselines like LC-Rec and TIGER.
 
-| Task | Metric | TIGER | LC-Rec-8B | **OneRec-8B-Pro** |
-| :--- | :--- | :--- | :--- | :--- |
-| **Short Video Rec** | Recall@32 | 0.0132 | 0.0180 | **0.0369** |
-| **Ad Rec** | Recall@32 | 0.0581 | 0.0723 | **0.0964** |
-| **Product Rec** | Recall@32 | 0.0283 | 0.0416 | **0.0538** |
-| **Label-Cond. Rec** | Recall@32 | 0.0123 | 0.0170 | **0.0235** |
-| **Label Pred.** | AUC | 0.6675 | 0.6139 | **0.6912** |
-| **Interactive Rec** | Recall@32 | -- | 0.2394 | **0.3458** |
-| **Item Understand.** | LLM-Judge Score| -- | 0.2517 | **0.3209** |
-| **Rec. Explanation** | LLM-Judge Score| -- | 3.9350 | **4.0381** |
-
+| Task | Metric | SASRec | TIGER | LC-Rec | OneRec-1.7B | OneRec-8B | OneRec-1.7B-Pro | **OneRec-8B-Pro** |
+| :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
+| **Short Video Rec** | Recall@32 | 0.0119 | 0.0132 | 0.0180 | 0.0272 | 0.0355 | 0.0274 | **0.0369** |
+| **Ad Rec** | Recall@32 | 0.0293 | 0.0581 | 0.0723 | 0.0707 | 0.0877 | 0.0735 | **0.0964** |
+| **Product Rec** | Recall@32 | 0.0175 | 0.0283 | 0.0416 | 0.0360 | 0.0470 | 0.0405 | **0.0538** |
+| **Label-Cond. Rec** | Recall@32 | 0.0140 | 0.0123 | 0.0170 | 0.0184 | 0.0228 | 0.0182 | **0.0235** |
+| **Label Pred.** | AUC | 0.6244 | 0.6675 | 0.6139 | 0.6184 | 0.6615 | 0.6071 | **0.6912** |
+| **Interactive Rec** | Recall@32 | -- | -- | 0.2394 | 0.1941 | 0.3032 | 0.2024 | **0.3458** |
+| **Item Und.** | LLM Score | -- | -- | 0.2517 | 0.3175 | 0.3202 | 0.3133 | **0.3209** |
+| **Rec. Explanation** | LLM Score | -- | -- | 3.9350 | 3.3540 | 3.6774 | 3.5060 | **4.0381** |
 
 ### Cross-Domain Transferability
 On the **Amazon Benchmark** (10 datasets), OpenOneRec demonstrates exceptional zero-shot/few-shot transfer capabilities, achieving an average **26.8% improvement** in Recall@10 over the second-best method.
 
+| Domain | SASRec | TIGER | LC-Rec | **Ours** |
+| :--- | :--- | :--- | :--- | :--- |
+| Baby | 0.0381 | 0.0318 | 0.0344 | **0.0513** |
+| Beauty | 0.0639 | 0.0628 | 0.0764 | **0.0924** |
+| Cell Phones | 0.0782 | 0.0786 | 0.0883 | **0.1036** |
+| Grocery | 0.0789 | 0.0691 | 0.0790 | **0.1029** |
+| Health | 0.0506 | 0.0534 | 0.0616 | **0.0768** |
+| Home | 0.0212 | 0.0216 | 0.0293 | **0.0390** |
+| Pet Supplies | 0.0607 | 0.0542 | 0.0612 | **0.0834** |
+| Sports | 0.0389 | 0.0331 | 0.0418 | **0.0547** |
+| Tools | 0.0437 | 0.0344 | 0.0438 | **0.0593** |
+| Toys | 0.0658 | 0.0527 | 0.0549 | **0.0953** |
+
+*Metric: Recall@10. Ours refers to OneRec-Foundation with text-augmented itemic tokens strategy.*
+
 ## 🚀 Quick Start
 
 *Code release and detailed usage instructions are coming soon.*
@@ -108,11 +120,9 @@ Currently, you can load our models using `transformers`:
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
-
 model_id = "Onerec/OneRec-Open-8B"
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
-
 # Example inference code will be updated here
 ```
 
@@ -128,4 +138,4 @@ If you find our work helpful, please cite our technical report:
 }
 ```
 ## 🛡️ License
-The code in this repository is licensed under the Apache 2.0 License. The model weights are subject to their specific license agreements.
+The code in this repository is licensed under the Apache 2.0 License. The model weights are subject to their specific license agreements.