We build open datasets, compact models, and tools for deep search agents — from evidence-grounded question construction to long-horizon reasoning and trustworthy evaluation.
4B / 9B
SFT & RL checkpoints
5,762
SFT trajectories · 4,448 unique questions
1,308
RL questions with reference answers
Start with Glasseek
MODEL · SUPERVISED FINE-TUNING
Glasseeker-4B-SFT ↗
A compact deep search checkpoint built on Qwen3.5-4B.
MODEL · SUPERVISED FINE-TUNING
Glasseeker-9B-SFT ↗
A 9B checkpoint for multi-step search planning and evidence synthesis.
MODEL · REINFORCEMENT LEARNING
Glasseeker-4B-RL ↗
The released reinforcement-learning checkpoint at the 4B scale.
MODEL · REINFORCEMENT LEARNING
Glasseeker-9B-RL ↗
The released reinforcement-learning checkpoint at the 9B scale.
DATASET · SEARCH TRAJECTORIES
Glasseek-SFT-5.7K ↗
5,762 trajectories covering 4,448 unique questions, with tool calls, observations, and answers.
DATASET · REINFORCEMENT LEARNING
Glasseek-RL-1.3K ↗
1,308 unique questions with reference answers for answer-based reward evaluation.
One connected research workflow
01 / CONSTRUCT
Verifiable data
Discover seeds, expand source-supported clues, and verify question uniqueness and solvability.
Data pipeline ↗02 / TRAIN
Compaction-aware RL
CM-GSPO optimizes execution sub-trajectories across context-compaction boundaries.
RL training code forthcoming03 / EVALUATE
Trustworthy evaluation
A Hermes-based harness with safeguards against benchmark-answer retrieval.
Evaluation runner ↗Earlier ICA resources
Our earlier releases remain available separately: ICA-SFT-14k and ICA-RL-Sample-400.