pravsels/lam-binpack-finetune

LAM Fine-Tuned on Bin-Pick-Pack

Fine-tuned DreamDojo LAM (Latent Action Model, 710M params) on the bin_pick_pack_coffee_capsules manipulation dataset.

Training Details

• Base model: LAM_400k.ckpt (pre-trained on GR1 humanoid data)
• Dataset: villekuosmanen/bin_pick_pack_coffee_capsules (42846 train pairs, 4819 val pairs)
• Resolution: 240x320
• Epochs: 57 completed (best at epoch 17, stopped early)
• Batch size: 32
• Learning rate: 1e-5
• Weight decay: 0.01
• KL beta: 1e-6
• Gradient clipping: 0.3
• Hardware: NVIDIA GH200 (Isambard HPC)
• Training time: ~12h

Results

Metric	Epoch 0	Epoch 17 (best)	Epoch 56 (final)
train_loss	0.000154	0.000076	0.000058
val_loss	0.000137	0.000097	0.000105
val_mse	0.000107	0.000080	0.000092
val_kl	29.35	16.57	12.83

Val loss improved until epoch 17 then plateaued around 1.0e-4. Train loss continued decreasing. Mild overfitting but no divergence.

Checkpoint

• File: best.pt (params only, 2.84 GB)
• Contents: model_state_dict, epoch, step, best_loss
• SHA-256: 72e746704080266c7c6aa265035de3bd2132b9ad2783dbfe8d9fc82670a838dc

Verify with:

sha256sum best.pt

Usage

import torch

ckpt = torch.load("best.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"])

Config

See lam_finetune_isambard.yaml for the full training configuration.

W&B

Training curves: wandb.ai/pravsels/lam-finetune/runs/afu3164m