You can deploy this model and start earning money today!

atrost/
nanochat-d24-matformer-ffnlog-fullattn-sft-l

Name: nanochat-d24-matformer-ffnlog-fullattn-sft-l
Availability: InStock

nanochat

extracted-submodel

sft

Views

No views yet

atrost/nanochat-d24-matformer-ffnlog-fullattn-sft-l

This is a standalone nanochat-native checkpoint extracted from atrost/nanochat-d24-matformer-ffnlog-fullattn-sft/d24_matformer_ffnlog_fullattn_sft at step 000485.

Family: matformer
Submodel: L
Config: {"attn_alpha_init_value": 1.0, "dec_alpha_init_value": 1.0, "ffn_alpha_init_value": 1.0, "head_dim": 80, "matformer_ffn_hidden_dims": [2560], "matformer_head_counts": [16], "n_embd": 1280, "n_head": 16, "n_kv_head": 16, "n_layer": 24, "norm_type": "rmsnorm", "sequence_len": 2048, "vocab_size": 32768, "window_pattern": "L"}

Load it into the local nanochat checkpoint layout before evaluation:

python

from nanochat.hf_utils import stage_hf_checkpoint

stage_hf_checkpoint(
    repo_id="atrost/nanochat-d24-matformer-ffnlog-fullattn-sft-l",
    path_in_repo="d24_matformer_ffnlog_fullattn_sft_l",
    model_tag="d24_matformer_ffnlog_fullattn_sft_l",
    step=485,
    base_dir="/tmp/nanochat-eval",
)

Then run eval with:

NANOCHAT_BASE_DIR=/tmp/nanochat-eval python -m scripts.base_eval --model-tag d24_matformer_ffnlog_fullattn_sft_l --step 485