Upload config/config.yaml with huggingface_hub
Browse files- config/config.yaml +48 -0
config/config.yaml
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
d_model: 768
|
| 2 |
+
n_layers: 12
|
| 3 |
+
vocab_size: 50257
|
| 4 |
+
attention:
|
| 5 |
+
type: mla
|
| 6 |
+
n_heads: 12
|
| 7 |
+
n_kv_heads: null
|
| 8 |
+
kv_compression_dim: 256
|
| 9 |
+
q_compression_dim: 384
|
| 10 |
+
rope_dim: 64
|
| 11 |
+
dropout: 0.0
|
| 12 |
+
ffn:
|
| 13 |
+
type: moe_swiglu
|
| 14 |
+
hidden_mult: 2.6667
|
| 15 |
+
n_experts: 4
|
| 16 |
+
top_k: 2
|
| 17 |
+
shared_experts: 1
|
| 18 |
+
load_balance_weight: 0.01
|
| 19 |
+
dropout: 0.0
|
| 20 |
+
position:
|
| 21 |
+
type: decoupled_rope
|
| 22 |
+
max_seq_len: 2048
|
| 23 |
+
rope_base: 10000.0
|
| 24 |
+
rope_dim: null
|
| 25 |
+
norm:
|
| 26 |
+
type: rmsnorm
|
| 27 |
+
eps: 1.0e-06
|
| 28 |
+
output:
|
| 29 |
+
type: tied
|
| 30 |
+
training:
|
| 31 |
+
batch_size: 8
|
| 32 |
+
seq_len: 1024
|
| 33 |
+
lr: 0.0003
|
| 34 |
+
min_lr: 3.0e-05
|
| 35 |
+
warmup_steps: 1000
|
| 36 |
+
max_steps: 50000
|
| 37 |
+
weight_decay: 0.1
|
| 38 |
+
grad_clip: 1.0
|
| 39 |
+
beta1: 0.9
|
| 40 |
+
beta2: 0.95
|
| 41 |
+
dataset: HuggingFaceFW/fineweb-edu
|
| 42 |
+
tokenizer: gpt2
|
| 43 |
+
log_interval: 10
|
| 44 |
+
eval_interval: 500
|
| 45 |
+
save_interval: 2500
|
| 46 |
+
eval_steps: 100
|
| 47 |
+
embed_dropout: 0.0
|
| 48 |
+
residual_dropout: 0.0
|