simplex-ai-inc/LiteResearcher-4B-SFT
LiteResearcher-4B-SFT by simplex-ai-inc is a 4 billion parameter SFT cold-start checkpoint based on Qwen3-4B-Thinking-2507, designed as the initial policy for a two-stage curriculum RL training framework for deep-research agents. It teaches the model a ReAct-style 'think' → 'search' → 'visit' → 'answer' loop and strict output contract. This model is primarily intended for reproducing RL training from scratch within the LiteResearcher framework, preserving long chain-of-thought behavior.
Loading preview...
LiteResearcher-4B-SFT: SFT Cold-Start Checkpoint
This model, developed by simplex-ai-inc, is the Supervised Fine-Tuning (SFT) cold-start checkpoint for the LiteResearcher project, an agentic RL training framework for deep-research agents. It serves as the initial policy for a two-stage curriculum Reinforcement Learning (RL) process, leading to the final simplex-ai-inc/LiteResearcher-4B model.
Key Characteristics & Training:
- Base Model: Built upon
Qwen/Qwen3-4B-Thinking-2507, leveraging its 4 billion parameters andQwen3ForCausalLMarchitecture. - Context Length: Supports a maximum sequence length of 64K tokens, with RoPE θ = 5,000,000 for extended context handling.
- Training Data: Fine-tuned on
simplex-ai-inc/LiteResearcher-Data, comprising approximately 68.2k SFT trajectories. - Core Functionality: Teaches the model a specific ReAct-style loop:
think→search→visit→answer, along with a strict<answer>...</answer>output contract essential for the RL environment. It preserves the long chain-of-thought behavior from its base model within<think>...</think>blocks.
Recommended Use Cases:
- RL Training Initialization: The primary use is as the starting point for reproducing the two-stage RL training process described in the LiteResearcher repository.
- Stand-alone Inference: Can be used for inference, expecting the LiteResearcher ReAct system prompt and tool schema. However, for a fully trained RL agent, the
simplex-ai-inc/LiteResearcher-4Bmodel is recommended.