Commit
•
659759f
1
Parent(s):
3c7fecb
Upload axolotl_config.yml
Browse files- axolotl_config.yml +83 -0
axolotl_config.yml
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: Qwen/Qwen1.5-0.5B
|
2 |
+
model_type: Qwen2ForCausalLM
|
3 |
+
tokenizer_type: AutoTokenizer
|
4 |
+
trust_remote_code: true
|
5 |
+
save_safetensors: true
|
6 |
+
|
7 |
+
load_in_8bit: false
|
8 |
+
load_in_4bit: false
|
9 |
+
strict: false
|
10 |
+
|
11 |
+
datasets:
|
12 |
+
- path: garage-bAInd/Open-Platypus
|
13 |
+
type: alpaca
|
14 |
+
prompt_style: chatml
|
15 |
+
- path: teknium/OpenHermes-2.5
|
16 |
+
type: sharegpt
|
17 |
+
conversation: qwen-7b-chat
|
18 |
+
- path: databricks/databricks-dolly-15k
|
19 |
+
type:
|
20 |
+
field_system: ""
|
21 |
+
field_instruction: instruction
|
22 |
+
field_input: context
|
23 |
+
field_output: response
|
24 |
+
format: |-
|
25 |
+
<|im_start|>system
|
26 |
+
You are a helpful assistant. Please give a concise and accurate answer<|im_end|>
|
27 |
+
<|im_start|>user
|
28 |
+
{instruction} {input}<|im_end|>
|
29 |
+
<|im_start|>assistant
|
30 |
+
no_input_format: |-
|
31 |
+
<|im_start|>system
|
32 |
+
You are a helpful assistant. Please give a concise and accurate answer<|im_end|>
|
33 |
+
<|im_start|>user
|
34 |
+
{instruction}<|im_end|>
|
35 |
+
<|im_start|>assistant
|
36 |
+
shuffle_merged_datasets: true
|
37 |
+
val_set_size: 0.04
|
38 |
+
chat_template: chatml
|
39 |
+
default_system_message: "You are a helpful assistant. Please give a concise and accurate answer"
|
40 |
+
output_dir: ./qwen_out
|
41 |
+
|
42 |
+
sequence_len: 2048
|
43 |
+
sample_packing: true
|
44 |
+
eval_sample_packing: false
|
45 |
+
pad_to_sequence_len: true
|
46 |
+
|
47 |
+
adapter: lora
|
48 |
+
lora_r: 8
|
49 |
+
lora_alpha: 16
|
50 |
+
lora_dropout: 0.05
|
51 |
+
lora_target_modules:
|
52 |
+
- q_proj
|
53 |
+
- v_proj
|
54 |
+
lora_target_linear: true
|
55 |
+
lora_modules_to_save:
|
56 |
+
- embed_tokens
|
57 |
+
- lm_head
|
58 |
+
|
59 |
+
wandb_project: qwen-0.5b-lora
|
60 |
+
wandb_name: qwen-lora
|
61 |
+
wandb_log_model: checkpoint
|
62 |
+
|
63 |
+
gradient_accumulation_steps: 16
|
64 |
+
micro_batch_size: 1
|
65 |
+
num_epochs: 4
|
66 |
+
optimizer: adamw_torch_fused
|
67 |
+
lr_scheduler: cosine
|
68 |
+
learning_rate: 0.0002
|
69 |
+
max_grad_norm: 1.0
|
70 |
+
|
71 |
+
train_on_inputs: false
|
72 |
+
group_by_length: false
|
73 |
+
bf16: true
|
74 |
+
|
75 |
+
gradient_checkpointing: false
|
76 |
+
logging_steps: 1
|
77 |
+
flash_attention: false
|
78 |
+
deepspeed: deepspeed_configs/zero1.json
|
79 |
+
|
80 |
+
warmup_steps: 4
|
81 |
+
evals_per_epoch: 0
|
82 |
+
saves_per_epoch: 1
|
83 |
+
weight_decay: 0.01
|