Downloads · 30 days
0
gemma-4-26B-A4B-DeepScaleR-medium-s42-fullckpt-step90
JWei05/gemma-4-26B-A4B-DeepScaleR-medium-s42-fullckpt-step90
gemma-4-26B-A4B-DeepScaleR-medium-s42-fullckpt-step90 is a machine learning model from JWei05. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as gemma.
DAPO/GRPO RL full checkpoint (verl FSDP2 fork). Not inference-only: it holds everything needed to resume training on another cluster.
Access
Public
Updated Sep 3, 2026
Repo size
359 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt305 GB · 85%
At a glance
- License
- gemma
- Access
- Public
- Created
- Sep 3, 2026
- Updated
- Sep 3, 2026
- SHA
- 97bde0df
Base models
Model card
From the Hugging Face model README
google/gemma-4-26B-A4B — DeepScaleR medium band — seed 42 — full resumable checkpoint (global step 90)
DAPO/GRPO RL full checkpoint (verl FSDP2 fork). Not inference-only: it holds everything needed to resume training on another cluster.
actor/model_world_size_8_rank_*.pt— FSDP2-sharded model weights.actor/optim_world_size_8_rank_*.pt— Adam optimizer state shards.actor/extra_state_world_size_8_rank_*.pt— RNG and LR-scheduler state.actor/huggingface/— tokenizer, config, and a consolidatedmodel.safetensors.data.pt— StatefulDataLoader cursor (exact dataset position).validation_early_stopping.json— early-stopping state.- Best mean@16 so far: 0.656 at step 90.
Resume
Sharded at world_size=8, so resume on 8 GPUs with the same FSDP
layout, keeping data paths, shuffle seed, batch size, and response count unchanged:
RESUME_MODE=resume_path trainer.resume_from_path=<local_dir>/global_step_90
- License
- gemma
- Created
- Sep 3, 2026
- Updated
- Sep 3, 2026