arl game24 multi-seed GRPO checkpoints

8 independent GRPO runs (seeds 0-7) finetuning Qwen/Qwen3-0.6B on game24, saved at training steps 1, 4, 8, 12, 200. Weights only (no optimizer state).

Layout

seed{N}/checkpoint-{K}/   # N in 0..7, K in 1, 4, 8, 12, 200

Each checkpoint is a standard HF model dir; load with:

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("Ksgk-fy/arl-game24-multiseed", subfolder="seed0/checkpoint-200")
tok = AutoTokenizer.from_pretrained("Ksgk-fy/arl-game24-multiseed", subfolder="seed0/checkpoint-200")

These accompany a study of the lottery gap (union vs. average accuracy across seeds), checkpoint divergence in weight space, and model merging / souping. checkpoint-200 is the final model; checkpoint-1/4/8/12 trace the early trajectory.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Ksgk-fy/arl-game24-multiseed

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Finetuned

(1035)

this model