---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3-8B
pipeline_tag: text-generation
tags:
- agent
- alfworld
- reinforcement-learning
- qwen3
---

# Qwen3-8B ALFWorld RL (step 570)

Qwen3-8B fine-tuned with reinforcement learning on the ALFWorld text-world
benchmark. This is a snapshot at training step 570.

* **Base model:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)
* **Task:** ALFWorld (TextWorld variant; 6 task types, valid_seen / valid_unseen)
* **Training step:** 570
* **Validation success rate (valid_seen, T=0.4, 8 rollouts/game):** ~0.957
* **Inference:** Qwen3 chat template with thinking enabled (`enable_thinking=True`).
  The model reasons inside `<think>...</think>` and emits the chosen action
  inside `<action>...</action>` tags, conditioned on the current observation
  and admissible-action list.
* **Tokenizer:** identical to `Qwen/Qwen3-8B` (vocab 151643).

## Intended use

Distillation source for smaller students (Qwen3-0.6B / Qwen2.5-0.5B), and
as a strong baseline for ALFWorld policy work. Released alongside the
training recipe at <https://github.com/MiaoLu3/verl>.