atrost/climbmix-stairformer-353m-1p2b-h100

ClimbMix StairFormer 353M

StairFormer trained on 1.2B tokens from nvidia/Nemotron-ClimbMix with nested auxiliary loss.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "atrost/climbmix-stairformer-353m-1p2b-h100"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    trust_remote_code=True,
    torch_dtype="auto",
)

trust_remote_code=True is required for the custom StairFormer/asymmetric checkpoints and harmless for the dense Llama baseline.

Source revision

This clean repo was split out from atrost/climbmix-llama-288m-2p8b-h100 at commit a5b6981.