andysalerno
commited on
Commit
•
37292fc
1
Parent(s):
45ba067
Create axolotl.yml
Browse files- axolotl.yml +90 -0
axolotl.yml
ADDED
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: ./mistralai/Mistral-7B-v0.1-chatml
|
2 |
+
model_type: MistralForCausalLM
|
3 |
+
tokenizer_type: LlamaTokenizer
|
4 |
+
is_mistral_derived_model: true
|
5 |
+
|
6 |
+
load_in_8bit: false
|
7 |
+
load_in_4bit: true
|
8 |
+
strict: false
|
9 |
+
|
10 |
+
datasets:
|
11 |
+
- path: andysalerno/ansalern-nectar-inputoutput
|
12 |
+
type:
|
13 |
+
field_instruction: input
|
14 |
+
field_output: output
|
15 |
+
format: "{instruction}"
|
16 |
+
no_input_format: "{instruction}"
|
17 |
+
dataset_prepared_path: last_run_prepared
|
18 |
+
val_set_size: 0.005
|
19 |
+
output_dir: ./qlora-out
|
20 |
+
|
21 |
+
adapter: qlora
|
22 |
+
lora_model_dir:
|
23 |
+
|
24 |
+
sequence_len: 8192
|
25 |
+
sample_packing: true # was true
|
26 |
+
eval_sample_packing: false
|
27 |
+
pad_to_sequence_len: true
|
28 |
+
|
29 |
+
lora_r: 32
|
30 |
+
lora_alpha: 16
|
31 |
+
lora_dropout: 0.05
|
32 |
+
lora_target_linear: true
|
33 |
+
lora_fan_in_fan_out:
|
34 |
+
lora_modules_to_save: ['embed_tokens', 'lm_head']
|
35 |
+
lora_target_modules:
|
36 |
+
- gate_proj
|
37 |
+
- down_proj
|
38 |
+
- up_proj
|
39 |
+
- q_proj
|
40 |
+
- v_proj
|
41 |
+
- k_proj
|
42 |
+
- o_proj
|
43 |
+
- embed_tokens
|
44 |
+
- lm_head
|
45 |
+
|
46 |
+
wandb_project: axolotl
|
47 |
+
wandb_entity:
|
48 |
+
wandb_watch:
|
49 |
+
wandb_name:
|
50 |
+
wandb_log_model:
|
51 |
+
|
52 |
+
gradient_accumulation_steps: 4
|
53 |
+
micro_batch_size: 2
|
54 |
+
num_epochs: 1
|
55 |
+
optimizer: adamw_bnb_8bit
|
56 |
+
lr_scheduler: cosine
|
57 |
+
learning_rate: 0.00005
|
58 |
+
|
59 |
+
train_on_inputs: false
|
60 |
+
group_by_length: false
|
61 |
+
bf16: auto
|
62 |
+
fp16:
|
63 |
+
tf32: false
|
64 |
+
|
65 |
+
gradient_checkpointing: true
|
66 |
+
early_stopping_patience: 3
|
67 |
+
resume_from_checkpoint:
|
68 |
+
local_rank:
|
69 |
+
logging_steps: 1
|
70 |
+
xformers_attention:
|
71 |
+
flash_attention: true
|
72 |
+
|
73 |
+
loss_watchdog_threshold: 5.0
|
74 |
+
loss_watchdog_patience: 3
|
75 |
+
|
76 |
+
warmup_steps: 0.1
|
77 |
+
eval_steps: 50
|
78 |
+
eval_table_size:
|
79 |
+
eval_table_max_new_tokens: 128
|
80 |
+
save_steps: 300
|
81 |
+
max_steps: 300
|
82 |
+
debug:
|
83 |
+
deepspeed:
|
84 |
+
weight_decay: 0.0
|
85 |
+
fsdp:
|
86 |
+
fsdp_config:
|
87 |
+
special_tokens:
|
88 |
+
bos_token: "<|im_start|>"
|
89 |
+
eos_token: "<|im_end|>"
|
90 |
+
unk_token: "<unk>"
|