LongTraceRL
Collection
LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards • 4 items • Updated • 1
LongTraceRL-8B
Model Description
LongTraceRL-8B is an 8-billion parameter reasoning model trained with reinforcement learning on long-context multi-hop QA tasks using trajectory-based tiered distractors and entity-level rubric rewards.
Model Details
- Base Model: DeepSeek-R1-0528-Qwen3-8B
- Parameters: 8B
- Architecture: Qwen3 (36 layers, hidden size 4096, GQA with 8 KV groups)
- Training Method: GRPO with entity-level rubric reward
- Context Length: 128K prompt + 32K response
- Language: English
Training Details
- Training Data: 2,815 long-context multi-hop QA samples (LongTraceRL Dataset)
- Training Steps: 200
- Learning Rate: 2e-6 (constant)
- Global Batch Size: 128
- GRPO Group Size: 8
- Rubric Reward Weight (η): 0.3
- Framework: Slime (Megatron-LM + SGLang)
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("THU-KEG/LongTraceRL-8B")
tokenizer = AutoTokenizer.from_pretrained("THU-KEG/LongTraceRL-8B")
Citation
Model tree for THU-KEG/LongTraceRL-8B
Base model
deepseek-ai/DeepSeek-R1-0528-Qwen3-8B