update config embeddings
Browse files- config.json +1 -1
- configs/minotaur.yml +107 -0
config.json
CHANGED
@@ -11,7 +11,7 @@
|
|
11 |
"initializer_range": 0.02,
|
12 |
"intermediate_size": 16384,
|
13 |
"layer_norm_eps": 1e-05,
|
14 |
-
"max_position_embeddings":
|
15 |
"model_type": "gpt_neox",
|
16 |
"num_attention_heads": 32,
|
17 |
"num_hidden_layers": 32,
|
|
|
11 |
"initializer_range": 0.02,
|
12 |
"intermediate_size": 16384,
|
13 |
"layer_norm_eps": 1e-05,
|
14 |
+
"max_position_embeddings": 8192,
|
15 |
"model_type": "gpt_neox",
|
16 |
"num_attention_heads": 32,
|
17 |
"num_hidden_layers": 32,
|
configs/minotaur.yml
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: /workspace/pythia-6.9b-deduped-8k/
|
2 |
+
base_model_config: EleutherAI/pythia-6.9b-deduped
|
3 |
+
model_type: GPTNeoXForCausalLM
|
4 |
+
tokenizer_type: AutoTokenizer
|
5 |
+
load_in_8bit: false
|
6 |
+
load_in_4bit: false
|
7 |
+
strict: False
|
8 |
+
gptq: false
|
9 |
+
device_map: auto
|
10 |
+
push_dataset_to_hub: winglian
|
11 |
+
hf_use_auth_token: true
|
12 |
+
# dataset_shard_num: 3
|
13 |
+
# dataset_shard_idx: 0
|
14 |
+
datasets:
|
15 |
+
- path: winglian/pygmalion-cleaned
|
16 |
+
data_files:
|
17 |
+
- v13_no_ai.cleaned.jsonl
|
18 |
+
type: pygmalion
|
19 |
+
- path: winglian/evals
|
20 |
+
data_files:
|
21 |
+
- hf/ARC-Challenge.jsonl
|
22 |
+
- hf/ARC-Easy.jsonl
|
23 |
+
- hf/riddle_sense.jsonl
|
24 |
+
type: explainchoice:chat
|
25 |
+
- path: winglian/evals
|
26 |
+
data_files:
|
27 |
+
- openai/tldr.jsonl
|
28 |
+
type: summarizetldr:chat
|
29 |
+
- path: winglian/evals
|
30 |
+
data_files:
|
31 |
+
- hf/gsm8k.jsonl
|
32 |
+
type: alpaca_chat.load_qa
|
33 |
+
- path: winglian/evals
|
34 |
+
data_files:
|
35 |
+
- hellaswag/hellaswag.jsonl
|
36 |
+
type: explainchoice:chat
|
37 |
+
- path: metaeval/ScienceQA_text_only
|
38 |
+
type: concisechoice:chat
|
39 |
+
- path: ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered
|
40 |
+
type: alpaca_chat
|
41 |
+
- path: ehartford/wizard_vicuna_70k_unfiltered
|
42 |
+
type: sharegpt:chat
|
43 |
+
- path: winglian/chatlogs-en-cleaned
|
44 |
+
data_files:
|
45 |
+
- sharegpt_cleaned.jsonl
|
46 |
+
type: sharegpt:chat
|
47 |
+
- path: teknium/GPT4-LLM-Cleaned
|
48 |
+
type: alpaca_chat
|
49 |
+
- path: teknium/GPTeacher-General-Instruct
|
50 |
+
data_files: gpt4-instruct-similarity-0.6-dataset.json
|
51 |
+
type: gpteacher:chat
|
52 |
+
- path: ewof/code-alpaca-instruct-unfiltered
|
53 |
+
type: alpaca_chat
|
54 |
+
- path: QingyiSi/Alpaca-CoT
|
55 |
+
data_files:
|
56 |
+
- Chain-of-Thought/formatted_cot_data/aqua_train.json
|
57 |
+
- Chain-of-Thought/formatted_cot_data/creak_train.json
|
58 |
+
- Chain-of-Thought/formatted_cot_data/ecqa_train.json
|
59 |
+
- Chain-of-Thought/formatted_cot_data/esnli_train.json
|
60 |
+
- Chain-of-Thought/formatted_cot_data/gsm8k_train.json
|
61 |
+
- Chain-of-Thought/formatted_cot_data/qasc_train.json
|
62 |
+
- Chain-of-Thought/formatted_cot_data/qed_train.json
|
63 |
+
- Chain-of-Thought/formatted_cot_data/sensemaking_train.json
|
64 |
+
- Chain-of-Thought/formatted_cot_data/strategyqa_train.json
|
65 |
+
- GPTeacher/Roleplay/formatted_roleplay-similarity_0.6-instruct-dataset.json
|
66 |
+
type: alpaca_chat
|
67 |
+
dataset_prepared_path: last_run_prepared
|
68 |
+
val_set_size: 0.01
|
69 |
+
adapter:
|
70 |
+
lora_model_dir:
|
71 |
+
sequence_len: 8192
|
72 |
+
max_packed_sequence_len: 8192
|
73 |
+
lora_r: 64
|
74 |
+
lora_alpha: 32
|
75 |
+
lora_dropout: 0.0
|
76 |
+
lora_target_modules:
|
77 |
+
lora_target_linear: true
|
78 |
+
lora_fan_in_fan_out: true # pythia/GPTNeoX lora specific
|
79 |
+
wandb_project: minotaur-7b
|
80 |
+
wandb_watch:
|
81 |
+
wandb_run_id:
|
82 |
+
wandb_log_model:
|
83 |
+
output_dir: ./minotaur-7b
|
84 |
+
gradient_accumulation_steps: 1
|
85 |
+
micro_batch_size: 4
|
86 |
+
num_epochs: 1
|
87 |
+
learning_rate: 0.00003
|
88 |
+
optimizer: adamw_bnb_8bit
|
89 |
+
lr_scheduler: cosine
|
90 |
+
warmup_steps: 100
|
91 |
+
save_steps: 84
|
92 |
+
eval_steps: 10
|
93 |
+
train_on_inputs: false
|
94 |
+
group_by_length: true
|
95 |
+
bf16: false
|
96 |
+
fp16: false
|
97 |
+
bfloat16: true
|
98 |
+
tf32: true
|
99 |
+
flash_optimum: true
|
100 |
+
early_stopping_patience:
|
101 |
+
resume_from_checkpoint:
|
102 |
+
local_rank:
|
103 |
+
gradient_checkpointing: true
|
104 |
+
fsdp:
|
105 |
+
fsdp_transformer_layer_cls_to_wrap:
|
106 |
+
collator_pad_to_longest: true
|
107 |
+
save_steps: 35
|