update readme and add exp logs

README.md

@@ -1,16 +1,27 @@
 # Search-R1: Train your LLMs to reason and call a search engine with reinforcement learning
 
-<strong>Search-R1</strong> is an extension of <strong>DeepSeek-R1(-Zero)</strong> methods for <em>training reasoning and searching (tool-call) interleaved LLMs</em>. We built upon [veRL](https://github.com/volcengine/verl).
+<!-- <strong>Search-R1</strong> is a reinforcement learning framework for <em>training reasoning and searching (tool-call) interleaved LLMs</em>.  -->
+<!-- We built upon [veRL](https://github.com/volcengine/verl). -->
+**Search-R1** is a reinforcement learning framework designed for training **reasoning-and-searching interleaved LLMs**—language models that learn to reason and make tool calls (e.g., to search engines) in a coordinated manner.
 
-Through RL (rule-based outcome reward), the 3B **base** LLM (both Qwen2.5-3b-base and Llama3.2-3b-base) develops reasoning and search engine calling abilities all on its own.
+<!-- It can be seen as an extension of <strong>DeepSeek-R1(-Zero)</strong> with interleaved search engine calling and an opensource RL training-based solution for <strong>OpenAI DeepResearch</strong>. -->
+Built upon [veRL](https://github.com/volcengine/verl), Search-R1 extends the ideas of **DeepSeek-R1(-Zero)** by incorporating interleaved search engine access and provides a fully open-source RL training pipeline. It serves as an alternative and open solution to **OpenAI DeepResearch**, enabling research and development in tool-augmented LLM reasoning.
 
-Paper: [link](https://arxiv.org/pdf/2503.09516); Model and data: [link](https://huggingface.co/collections/PeterJinGo/search-r1-67d1a021202731cb065740f5); Twitter thread: [link](https://x.com/BowenJin13/status/1895544294473109889); Full experiment log 1: [link](https://wandb.ai/peterjin/Search-R1-open); Full experiment log 2: [link](https://wandb.ai/peterjin/Search-R1-nq_hotpotqa_train)
+<!-- Through RL (rule-based outcome reward), the 3B **base** LLM (both Qwen2.5-3b-base and Llama3.2-3b-base) develops reasoning and search engine calling abilities all on its own. -->
 
-You can refer to this [link](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa) for detailed instructions on reproducing the results from the paper.
+We support different RL methods (e.g., PPO, GRPO, reinforce), different LLMs (e.g., llama3, Qwen2.5, etc) and different search engines (e.g., local sparse/dense retrievers and online search engines).
+
+Paper: [link](https://arxiv.org/pdf/2503.09516); Model and data: [link](https://huggingface.co/collections/PeterJinGo/search-r1-67d1a021202731cb065740f5); Twitter thread: [link](https://x.com/BowenJin13/status/1895544294473109889); Full experiment log: [prelim](https://wandb.ai/peterjin/Search-R1-open); [v0.1](https://wandb.ai/peterjin/Search-R1-nq_hotpotqa_train); [v0.2](https://wandb.ai/peterjin/Search-R1-v0.2). Details about these logs and methods can be find [here](https://github.com/PeterGriffinJin/Search-R1/blob/main/docs/experiment_log.md).
 
 
-![single-turn](public/single-turn.png)
+![single-turn](public/main.png)
 
+## News
+
+- [2025.4.10] We support [multinode](https://github.com/PeterGriffinJin/Search-R1/blob/main/docs/multinode.md) training for 30B+ LLMs!
+- [2025.4.5] We support [different search engines](https://github.com/PeterGriffinJin/Search-R1/blob/main/docs/retriever.md) including sparse local retriever, dense local retriever with ANN indexing and online search engines!
+- [2025.3] The [paper](https://arxiv.org/pdf/2503.09516) is published with the logs: [v0.1](https://wandb.ai/peterjin/Search-R1-nq_hotpotqa_train); [v0.2](https://wandb.ai/peterjin/Search-R1-v0.2).
+- [2025.2] We opensource Search-R1 codebase with [preliminary results](https://wandb.ai/peterjin/Search-R1-open).
 
 ## Links
 

docs/experiment_log.md

@@ -0,0 +1,33 @@
+
+## Experiment log
+
+### Preliminary results 
+
+Resources: [wandb](https://wandb.ai/peterjin/Search-R1-open)
+
+
+The preliminary experiment is conducted only on natural question (NQ) dataset (+ PPO) with a small number of training steps.
+
+
+### v0.1
+
+Resources: [wandb](https://wandb.ai/peterjin/Search-R1-nq_hotpotqa_train), [docs](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa), [scripts](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa/v0.1)
+
+
+We extend the experiments from NQ to seven datasets with both PPO and GRPO methods. The studies are still on a small number of training steps with a big learning rate warm up ratio.
+
+
+### v0.2
+
+Resources: [wandb](https://wandb.ai/peterjin/Search-R1-v0.2), [docs](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa), [scripts](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa/v0.2)
+
+
+We fix several bugs including [retrieved token masking](https://github.com/PeterGriffinJin/Search-R1/pull/21) and [GRPO sample indexing](https://github.com/PeterGriffinJin/Search-R1/commit/9ec2fa9892fbf0315d0c67b4dc08ae8f6cf5f378). 
+The former can largely improve the stablity of RL training. 
+Then we adjust the training scripts, increasing the number of training steps and decreasing the learning rate warm up ratio, to obtain a better performance, and conduct experiments on different scale of LLMs (3B, 7B, 14B).
+
+
+### v0.3
+
+Ongoing, stay tuned!
+