Add project page link

README.md

@@ -1,18 +1,16 @@
 ---
-license: apache-2.0
+base_model:
+- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
 datasets:
 - open-thoughts/OpenThoughts2-1M
 - Vinnnf/Hybrid-OpenThoughts2-1M-1.5B
-base_model:
-- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
-pipeline_tag: text-generation
 library_name: transformers
+license: apache-2.0
+pipeline_tag: text-generation
 ---
 
-
 # Thinkless: LLM Learns When to Think
 
-
 ![image/png](https://cdn-uploads.huggingface.co/production/uploads/646a1939c37ca1e12308fe81/SRxJKkSuC0y-oMB7SFeR6.png)
 
 <table>
@@ -47,6 +45,10 @@ library_name: transformers
       <td>📊 <strong>Data for RL</strong></td>
       <td><a href="https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset">agentica-org/DeepScaleR-Preview-Dataset</a></td>
     </tr>
+    <tr>
+      <td> 🌐 <strong>Project Page</strong></td>
+      <td><a href="https://sites.google.com/view/eagle-llm">Thinkless Website</a></td>
+    </tr>
   </tbody>
 </table>
 
@@ -55,7 +57,7 @@ library_name: transformers
 > [!NOTE] 
 > ***Can LLMs learn when to think?***
 
-We propose Thinkless, a learnable framework that empowers an LLM to adaptively select between short-form and long-form reasoning based on both task complexity and the model's ability. Thinkless is trained under a reinforcement learning paradigm and employs two control tokens, \<short\> for concise responses and \<think\> for detailed reasoning. At the core of our method is a Decoupled Group Relative Policy Optimization (DeGRPO) algorithm, which decomposes the learning objective of hybrid reasoning into two components: (1) a control token loss that governs the selection of the reasoning mode, and (2) a response loss that improves the accuracy of the generated answers. This decoupled formulation enables fine-grained control over the contributions of each objective, stabilizing training and effectively preventing collapse observed in vanilla GRPO. Empirically, on several benchmarks such as Minerva Algebra, MATH-500, and GSM8K, Thinkless is able to reduce the usage of long-chain thinking by 50\% - 90\%, significantly reducing the computational cost of Reasoning Language Models.
+We propose Thinkless, a learnable framework that empowers an LLM to adaptively select between short-form and long-form reasoning based on both task complexity and the model's ability. Thinkless is trained under a reinforcement learning paradigm and employs two control tokens, \<short\> for concise responses and \<think\> for detailed reasoning. At the core of our method is a Decoupled Group Relative Policy Optimization (DeGRPO) algorithm, which decomposes the learning objective of hybrid reasoning into two components: (1) a control token loss that governs the selection of the reasoning mode, and (2) a response loss that improves the accuracy of the generated answers. This decoupled formulation enables fine-grained control over the contributions of each objective, stabilizing training and effectively preventing collapse observed in vanilla GRPO. Empirically, on several benchmarks such as Minerva Algebra, MATH-500, and GSM8K, Thinkless is able to reduce the usage of long-chain thinking by 50% - 90%, significantly reducing the computational cost of Reasoning Language Models.
 
 
 ## Pipeline
@@ -76,7 +78,8 @@ model = AutoModelForCausalLM.from_pretrained(
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 
 instruction = "Please reason step by step, and put your final answer within \\boxed{}."
-prompt = f"{instruction}\nThe arithmetic mean of 7, 2, $x$ and 10 is 9. What is the value of $x$?"
+prompt = f"{instruction}
+The arithmetic mean of 7, 2, $x$ and 10 is 9. What is the value of $x$?"
 
 messages = [
     {"role": "user", "content": prompt}
@@ -108,7 +111,8 @@ num_tokens = len(generated_ids[0])
 response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
 
 print(text+response)
-print(f"\nThink Mode: {think_mode}")
+print(f"
+Think Mode: {think_mode}")
 print(f"Number of tokens: {num_tokens}")
 ```