Upload folder using huggingface_hub
Browse files- configs/tiny-mistral.yml +70 -0
configs/tiny-mistral.yml
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: openaccess-ai-collective/tiny-mistral
|
2 |
+
base_model_config: openaccess-ai-collective/tiny-mistral
|
3 |
+
model_type: MistralForCausalLM
|
4 |
+
tokenizer_type: LlamaTokenizer
|
5 |
+
is_mistral_derived_model: true
|
6 |
+
model_config:
|
7 |
+
dropout_p: 0.1
|
8 |
+
|
9 |
+
load_in_8bit: false
|
10 |
+
load_in_4bit: false
|
11 |
+
strict: false
|
12 |
+
|
13 |
+
datasets:
|
14 |
+
- path: emrgnt-cmplxty/sciphi-textbooks-are-all-you-need
|
15 |
+
type: completion
|
16 |
+
field: completion
|
17 |
+
|
18 |
+
dataset_prepared_path: last_run_prepared
|
19 |
+
val_set_size: 0.001
|
20 |
+
output_dir: ./mistral-textbooks
|
21 |
+
hub_model_id: winglian/tiny-mistral-textbooks
|
22 |
+
|
23 |
+
sequence_len: 4096
|
24 |
+
sample_packing: true
|
25 |
+
pad_to_sequence_len: true
|
26 |
+
|
27 |
+
wandb_project: tiny-mistral-textbooks
|
28 |
+
wandb_entity:
|
29 |
+
wandb_watch:
|
30 |
+
wandb_run_id:
|
31 |
+
wandb_log_model:
|
32 |
+
|
33 |
+
gradient_accumulation_steps: 8
|
34 |
+
micro_batch_size: 4
|
35 |
+
num_epochs: 1
|
36 |
+
optimizer: adamw_torch
|
37 |
+
lr_scheduler: cosine
|
38 |
+
adam_beta2: 0.95
|
39 |
+
adam_epsilon: 0.00001
|
40 |
+
max_grad_norm: 1.0
|
41 |
+
learning_rate: 0.000025
|
42 |
+
|
43 |
+
train_on_inputs: false
|
44 |
+
group_by_length: false
|
45 |
+
bf16: true
|
46 |
+
fp16: false
|
47 |
+
tf32: false
|
48 |
+
|
49 |
+
gradient_checkpointing: true
|
50 |
+
early_stopping_patience:
|
51 |
+
resume_from_checkpoint:
|
52 |
+
local_rank:
|
53 |
+
logging_steps: 1
|
54 |
+
xformers_attention:
|
55 |
+
flash_attention: true
|
56 |
+
|
57 |
+
warmup_steps: 10
|
58 |
+
eval_steps: 0.1
|
59 |
+
eval_table_size:
|
60 |
+
eval_table_max_new_tokens:
|
61 |
+
save_steps: 0.1
|
62 |
+
debug:
|
63 |
+
deepspeed:
|
64 |
+
weight_decay: 0.1
|
65 |
+
fsdp:
|
66 |
+
fsdp_config:
|
67 |
+
special_tokens:
|
68 |
+
bos_token: "<s>"
|
69 |
+
eos_token: "</s>"
|
70 |
+
unk_token: "<unk>"
|