Upload config.yaml with huggingface_hub
Browse files- config.yaml +54 -0
config.yaml
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: v4_moe_shared
|
| 2 |
+
ckpt_dir: checkpoints
|
| 3 |
+
model:
|
| 4 |
+
vocab_size: 32000
|
| 5 |
+
n_ctx: 256
|
| 6 |
+
d_model: 512
|
| 7 |
+
n_layers: 8
|
| 8 |
+
n_heads: 8
|
| 9 |
+
d_ff: 2048
|
| 10 |
+
dropout: 0.0
|
| 11 |
+
tie_embeddings: true
|
| 12 |
+
init_std: 0.02
|
| 13 |
+
languages:
|
| 14 |
+
- vi
|
| 15 |
+
- ja
|
| 16 |
+
ffn_variant: moe_shared
|
| 17 |
+
n_experts: 4
|
| 18 |
+
n_active_experts: 2
|
| 19 |
+
n_shared_experts: 1
|
| 20 |
+
d_ff_expert: null
|
| 21 |
+
load_balance_coef: 0.01
|
| 22 |
+
data:
|
| 23 |
+
dataset_id: belumind/en-vi-ja-curated-500k-triplets
|
| 24 |
+
tokenizer_path: tokenizers/a2_vi_ja_en.json
|
| 25 |
+
vocab_size: 32000
|
| 26 |
+
n_ctx: 256
|
| 27 |
+
languages:
|
| 28 |
+
- vi
|
| 29 |
+
- ja
|
| 30 |
+
seed: 0
|
| 31 |
+
token_budget: 73000000
|
| 32 |
+
train:
|
| 33 |
+
batch_rows: 32
|
| 34 |
+
grad_accum: 1
|
| 35 |
+
lr: 0.001
|
| 36 |
+
min_lr_ratio: 0.1
|
| 37 |
+
warmup_steps: 300
|
| 38 |
+
weight_decay: 0.1
|
| 39 |
+
betas:
|
| 40 |
+
- 0.9
|
| 41 |
+
- 0.95
|
| 42 |
+
grad_clip: 1.0
|
| 43 |
+
amp_dtype: auto
|
| 44 |
+
log_every: 50
|
| 45 |
+
eval_every: 500
|
| 46 |
+
eval_examples: 4000
|
| 47 |
+
ckpt_every: 2000
|
| 48 |
+
seed: 0
|
| 49 |
+
eval:
|
| 50 |
+
batch_rows: 64
|
| 51 |
+
max_new_tokens: 128
|
| 52 |
+
gen_batch_size: 256
|
| 53 |
+
bleu_samples: 10000
|
| 54 |
+
bleu_seed: 0
|