GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jjalora
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
jjalora
Model Tree
Input Modalities
Output Modalities
Supported Functionality
pythia, seed 11One from-scratch Pythia-410M pretraining run from the DIET weight-space campaign. The recipe
is the open-source Pythia recipe (EleutherAI/pythia models/410M/pythia-410m.yml): same
architecture, same token stream and order (EleutherAI/pile-standard-pythia-preshuffled,
batch i = rows [i·1024, (i+1)·1024)), AdamW lr 3e-4, betas (0.9, 0.95), weight decay 0.1
(no decay on LayerNorm/bias), grad clip 1.0, cosine to 3e-5 over 143,000 steps, 1,430 warmup
steps, fp16 autocast with fp32 master weights, world batch 1024 × 2048 tokens on every GPU
count. Only the random seed and the weight-initialization scheme differ between runs.
init_scheme = pythia — pythia is GPT-NeoX's small_init + wang_init; other schemes
are documented in src/init_schemes.py of the training code.stepN for N in {0, 1, 2, 4, …, 512} ∪ {every 100 steps}, weights
only, float16 safetensors. main holds the final weights once the run completes.pythia scheme matches the per-tensor weight distribution of
EleutherAI/pythia-410m revision step0 (checked in tests/test_pythia_step0_init_match.py).Provenance (campaign.json): {"cluster": "wheat", "data": "EleutherAI/pile-standard-pythia-preshuffled", "git_commit": "06ae800", "gpus_per_node": 4, "hardware": "A100-PCIE-40GB", "jobs": [{"job": "383968", "start": "2026-10-23 18:48:23"}], "micro_batch": 4, "nodes": 8, "recipe": "EleutherAI/pythia models/410M/pythia-410m.yml", "scheme": "pythia", "seed": 11, "world_batch": 1024}