Upload axolotl_config_lu170wzt.yml
Browse files- axolotl_config_lu170wzt.yml +143 -0
axolotl_config_lu170wzt.yml
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## model
|
2 |
+
base_model: /home/dgxuser/workspace/Mango/models/Vulpecula
|
3 |
+
model_type: AutoModelForCausalLM
|
4 |
+
tokenizer_type: AutoTokenizer
|
5 |
+
|
6 |
+
## qlora
|
7 |
+
load_in_8bit: false
|
8 |
+
load_in_4bit: false
|
9 |
+
strict: false
|
10 |
+
## Lora
|
11 |
+
adapter: lora
|
12 |
+
lora_model_dir:
|
13 |
+
lora_r: 64
|
14 |
+
lora_alpha: 32
|
15 |
+
lora_dropout: 0.0
|
16 |
+
peft_use_rslora: true
|
17 |
+
lora_target_modules:
|
18 |
+
lora_mlp_kernel: true
|
19 |
+
lora_qkv_kernel: true
|
20 |
+
lora_o_kernel: true
|
21 |
+
lora_target_linear: true
|
22 |
+
#lora_target_modules:
|
23 |
+
# - gate_proj
|
24 |
+
# - down_proj
|
25 |
+
# - up_proj
|
26 |
+
# - q_proj
|
27 |
+
# - v_proj
|
28 |
+
# - k_proj
|
29 |
+
# - o_proj
|
30 |
+
## data
|
31 |
+
datasets:
|
32 |
+
- path: PocketDoc/Dans-Personamaxx-VN
|
33 |
+
type: dan-chat-advanced-llama3
|
34 |
+
- path: NewEden/LIMARP-Complexity
|
35 |
+
type: dan-chat-advanced-llama3
|
36 |
+
- path: NewEden/PIPPA-Mega-Filtered
|
37 |
+
type: dan-chat-advanced-llama3
|
38 |
+
- path: NewEden/OpenCAI-ShareGPT
|
39 |
+
type: dan-chat-advanced-llama3
|
40 |
+
- path: NewEden/Creative_Writing-Complexity
|
41 |
+
type: dan-chat-advanced-llama3
|
42 |
+
- path: NewEden/Light-Novels-Roleplay-Logs-Books-Oh-My-duplicate-turns-removed
|
43 |
+
type: dan-chat-advanced-llama3
|
44 |
+
- path: PocketDoc/Dans-Failuremaxx-Adventure-3
|
45 |
+
type: dan-chat-advanced-llama3
|
46 |
+
- path: NewEden/Books-V2-ShareGPT
|
47 |
+
type: dan-chat-advanced-llama3
|
48 |
+
- path: NewEden/Deepseek-V3-RP-Filtered
|
49 |
+
type: dan-chat-advanced-llama3
|
50 |
+
- path: NewEden/Final-Alpindale-LNs-ShareGPT
|
51 |
+
type: dan-chat-advanced-llama3
|
52 |
+
- path: NewEden/DeepseekRP-Filtered
|
53 |
+
type: dan-chat-advanced-llama3
|
54 |
+
- path: NewEden/RP-logs-V2-Experimental
|
55 |
+
type: dan-chat-advanced-llama3
|
56 |
+
- path: anthracite-org/kalo_opus_misc_240827
|
57 |
+
type: dan-chat-advanced-llama3
|
58 |
+
- path: anthracite-org/kalo_misc_part2
|
59 |
+
type: dan-chat-advanced-llama3
|
60 |
+
- path: NewEden/Storium-Prefixed-Clean
|
61 |
+
type: dan-chat-advanced-llama3
|
62 |
+
shuffle_merged_datasets: true
|
63 |
+
dataset_prepared_path: base-dataset_prepared
|
64 |
+
val_set_size: 0.0
|
65 |
+
output_dir: ./SFT-Vulpecula
|
66 |
+
|
67 |
+
## Liger + CCE
|
68 |
+
plugins:
|
69 |
+
- axolotl.integrations.liger.LigerPlugin
|
70 |
+
- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
|
71 |
+
liger_rope: true
|
72 |
+
liger_rms_norm: true
|
73 |
+
liger_layer_norm: true
|
74 |
+
liger_glu_activation: true
|
75 |
+
liger_fused_linear_cross_entropy: false
|
76 |
+
cut_cross_entropy: true
|
77 |
+
|
78 |
+
## CTX settings
|
79 |
+
sequence_len: 8192
|
80 |
+
sample_packing: true
|
81 |
+
eval_sample_packing: false
|
82 |
+
pad_to_sequence_len: true
|
83 |
+
|
84 |
+
## WandB
|
85 |
+
wandb_project: Francois
|
86 |
+
wandb_entity:
|
87 |
+
wandb_watch:
|
88 |
+
wandb_name: Big_boy
|
89 |
+
wandb_log_model:
|
90 |
+
|
91 |
+
## evals
|
92 |
+
#evals_per_epoch: 4
|
93 |
+
#eval_table_size:
|
94 |
+
#eval_max_new_tokens: 128
|
95 |
+
|
96 |
+
## hparams
|
97 |
+
gradient_accumulation_steps: 3
|
98 |
+
micro_batch_size: 3
|
99 |
+
num_epochs: 2
|
100 |
+
optimizer: adamw_bnb_8bit
|
101 |
+
lr_scheduler: cosine
|
102 |
+
learning_rate: 2e-5
|
103 |
+
warmup_steps: 50
|
104 |
+
weight_decay: 0.0025
|
105 |
+
## max grad norm
|
106 |
+
max_grad_norm: 1.0
|
107 |
+
|
108 |
+
train_on_inputs: false
|
109 |
+
group_by_length: false
|
110 |
+
bf16: auto
|
111 |
+
fp16:
|
112 |
+
tf32: false
|
113 |
+
|
114 |
+
gradient_checkpointing: true
|
115 |
+
early_stopping_patience:
|
116 |
+
resume_from_checkpoint:
|
117 |
+
local_rank:
|
118 |
+
logging_steps: 1
|
119 |
+
xformers_attention:
|
120 |
+
flash_attention: true
|
121 |
+
s2_attention:
|
122 |
+
saves_per_epoch: 2
|
123 |
+
debug:
|
124 |
+
deepspeed: ./deepspeed_configs/zero3_bf16.json
|
125 |
+
|
126 |
+
#fsdp:
|
127 |
+
# - full_shard
|
128 |
+
# - auto_wrap
|
129 |
+
#fsdp_config:
|
130 |
+
# fsdp_limit_all_gathers: true
|
131 |
+
# fsdp_sync_module_states: true
|
132 |
+
# fsdp_offload_params: true
|
133 |
+
# fsdp_activation_checkpointing: true
|
134 |
+
# fsdp_use_orig_params: false
|
135 |
+
# fsdp_cpu_ram_efficient_loading: true
|
136 |
+
# fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
|
137 |
+
# fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
|
138 |
+
# fsdp_state_dict_type: FULL_STATE_DICT
|
139 |
+
# fsdp_sharding_strategy: FULL_SHARD
|
140 |
+
special_tokens:
|
141 |
+
pad_token: <|finetune_right_pad_id|>
|
142 |
+
eos_token: <|eot_id|>
|
143 |
+
|