Downloads · 30 days
0
exp0005-valv2ft100k2gpu
jungwook2358/exp0005-valv2ft100k2gpu
exp0005-valv2ft100k2gpu is a machine learning model from jungwook2358. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
openarm1 목적지: /home/rlwrld/Code/jungwook/checkpoints/exp0005valv2ft100k2gpu
Access
Public
Updated Aug 17, 2026
Repo size
23.8 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors28.3 GB · 100%
At a glance
- Access
- Public
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026
- SHA
- 73b62ddc
Model card
From the Hugging Face model README
EXP-0005 valv2-ft(2GPU/100k) eval bundle
openarm1 목적지: /home/rlwrld/Code/jungwook/checkpoints/exp0005_valv2ft100k2gpu
생성 소스 (sjw_alinlab 클러스터, /rlwrld-unified-checkpoints/jungwook/...):
| 번들 | 소스 | 비고 |
|---|---|---|
tokenizer_ft100k/ | action_tokenizer/checkpoints_action_tokenizer/joint_soupv1_v4_..._ft_only_openarm_prq_valv2_2gpu_100k/checkpoint-100000 | --actlat-tokenizer-path 로 지정. optimizer/rng 는 제외 (eval 불필요). 원본 model.safetensors 는 0600 이라 복사 후 644 로 완화. |
vla_cotrain_30k/ | rldx1/egopi_prq_cotrain_pnpvalv2merged_bsz64_30000_vl1_sd0_valv2ft100k2gpu_actlat/egopi_prq_cotrain_pnpvalv2merged_v3-208bal_.../ (run root) | 학습 종료 시점 = step 30000 최종 저장본 (HF shard 3개 + config + experiment_cfg + processor). deepspeed checkpoint-30000/global_step30000 optimizer state 는 제외. |
vla_robotonly_30k/ | rldx1/egopi_prq_robotonly_bsz64_30000_vl1_sd0_valv2ft100k2gpu_actlat/egopi_prq_robotonly_v3-208bal_.../ (run root) | 동일 |
assets/egopi_prq_stats.json | /sjw_alinlab/home/junhyeong/RLDX-1-egopi/assets/egopi/egopi_prq_stats.json | --actlat-prq-stats-path (latent→action 복원에 필수) |
assets/z_stats.npz | actlat_cache/openarm_prq_valv2_2gpu_ft100k/z_stats.npz | 두 run 모두 STANDARDIZE=0 이라 eval 에 불필요. 참고용. |
학습 스크립트: sbatch_scripts/multiemb/v4_soupv1/recon_dino_bn64_l1_mse_naiveln_vae_embtok_finetune_only_openarm_prq_valv2_2gpu_100k.sh
→ Stage-2 RLDX-1/run_scripts/train/exp0005/egopi_actlat_valv2_{cotrain,robotonly}.sh
(체인이 TOK_CKPT=.../ft_..._2gpu_100k/checkpoint-100000, WANDB_PROJECT=egopi-actlat-valv2tok-2gpu100k 로 오버라이드)
eval 시 주의: --actlat-mode --actlat-embodiment-id openarm_prq,
--actlat-tokenizer-path <tokenizer_ft100k>, --actlat-prq-stats-path <assets/egopi_prq_stats.json>.
디렉토리 이름의 valv2ft200k 접미사는 스크립트에 하드코딩된 문자열일 뿐, 실제 토크나이저는 2GPU/100k ft 이다.
SHA256SUMS 로 전송 검증: cd <dest> && sha256sum -c SHA256SUMS
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026