namkoong-lab/LatentGym_Qwen3-8B_10episodes_SingleLatent_number_guessing

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The namkoong-lab/LatentGym_Qwen3-8B_10episodes_SingleLatent_number_guessing model is an 8 billion parameter Qwen3-8B variant, fine-tuned using the GRPO algorithm for single-latent training on the number_guessing environment. It is part of the LatentGym testbed, specifically optimized for tasks involving cumulative reward over 10 episodes. This model is designed for research into reinforcement learning from human feedback (RLHF) and latent space exploration within controlled environments.

Loading preview...

Model Overview

This model, namkoong-lab/LatentGym_Qwen3-8B_10episodes_SingleLatent_number_guessing, is an 8 billion parameter variant of the Qwen3-8B architecture. It has been specifically fine-tuned using the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm. The primary focus of this training was on a single-latent setup within the number_guessing environment, where the model was exposed to the set_of_3 latent during training.

Key Characteristics

  • Base Model: Qwen/Qwen3-8B, a powerful large language model.
  • Fine-tuning Algorithm: GRPO, indicating a reinforcement learning approach to optimization.
  • Environment: Optimized for the number_guessing environment, a controlled setting for evaluating latent space learning.
  • Latent Training: Trained exclusively on the set_of_3 latent within the number_guessing environment.
  • Episode Length: Configured for 10 episodes per trajectory (N=10), with a cumulative reward function (Σᵢ Gᵢ).
  • Training Hyperparameters: Utilizes AdamW optimizer with a learning rate of 5e-07, constant_with_warmup LR schedule, and specific KL coefficient (β=0.04) and clip range (ε=0.2) settings.

Intended Use Cases

This model is particularly suitable for:

  • Research in RLHF: Exploring the effects of GRPO fine-tuning on specific tasks.
  • Latent Space Analysis: Investigating how models learn and utilize latent representations in controlled environments.
  • Benchmarking: Serving as a component within the LatentGym testbed for comparative studies on latent learning.
  • Controlled Experimentation: Ideal for researchers studying model behavior under specific, constrained training conditions like the number_guessing task.

Finetunes
1 models