Skip to content
Other

hal-rock/qwen3-4b-maze-grpo-fft

Qwen3-4B model fine-tuned with reinforcement learning in a maze

2 min read

This model is a fully fine-tuned version of Qwen3-4B-Instruct-2507 (a 4-billion-parameter language model) using the Dr. GRPO algorithm (reinforcement learning method for language models). The experiment explores how reinforcement learning (RL (reinforcement learning)) changes the model's environment predictions in a custom emoji maze. The model outputs step predictions for navigating through tiles. It is designed for AI researchers and studying language model training dynamics.

Key features

  • Four checkpoints: trained to a peak at step 25 and to the final state at step 95 with different reward configurations.
  • Maze environment: a 100x100 grid where the agent must make moves across tiles with gold and mold.
  • Wind factor: the simulation includes a probability of random agent displacement.
  • Training method: Dr. GRPO (group relative policy optimization algorithm) with AdamW steps and fp32 weights under bf16.

What you can do

  • Explore behavior: test the model's ability to navigate character-based mazes.
  • Compare checkpoints: analyze the difference between the training peak at step 25 and the final state at step 95.
  • Run inference: load specific checkpoints (main or swap) via transformers to analyze predictions.

Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "hal-rock/qwen3-4b-maze-grpo-fft"
tok = AutoTokenizer.from_pretrained(repo, subfolder="main/step-25")
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="main/step-25", dtype=torch.bfloat16)
Read moreDetails
• Rewards: +20 for gold, -10 for mold, -0.1 per step.
• Training: 64 mazes with 64 rollouts per step, lr 3e-6.
• Limitations: instability after step 25, rise in action entropy, and drop in rewards by step 95.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.2 GB · ready files9.5 GB
  • FP8 / INT8recommendedweights 4.1 GB5.0 GB
  • 4-bitweights 2.3 GB3.0 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 3080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # hal-rock/qwen3-4b-maze-grpo-fft
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download hal-rock/qwen3-4b-maze-grpo-fft --local-dir ~/models/qwen3-4b-maze-grpo-fft
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to qwen3-4b-maze-grpo-fft
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-4B-Instruct-2507 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face

About the model

hal-rock/qwen3-4b-maze-grpo-fft
Author
hal-rock
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments