Commit
•
cf790b9
1
Parent(s):
f42fe16
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,111 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
+
wandb: https://wandb.ai/open-assistant/supervised-finetuning/runs/sytsyhrp
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
Datasets:
|
| 8 |
+
|
| 9 |
+
```
|
| 10 |
+
pretrain:
|
| 11 |
+
num_train_epochs: 1
|
| 12 |
+
weight_decay: 0.0
|
| 13 |
+
use_custom_sampler: true
|
| 14 |
+
sort_by_length: false
|
| 15 |
+
datasets:
|
| 16 |
+
- gpteacher_roleplay:
|
| 17 |
+
val_split: 0.05
|
| 18 |
+
- red_pajama:
|
| 19 |
+
fraction: 0.25
|
| 20 |
+
max_val_set: 1000
|
| 21 |
+
- wizardlm_70k:
|
| 22 |
+
val_split: 0.05
|
| 23 |
+
max_val_set: 500
|
| 24 |
+
- joke:
|
| 25 |
+
val_split: 0.05
|
| 26 |
+
- poem_instructions:
|
| 27 |
+
val_split: 0.025
|
| 28 |
+
- oa_stackexchange:
|
| 29 |
+
val_split: 0.05
|
| 30 |
+
fraction: 0.1
|
| 31 |
+
max_val_set: 1000
|
| 32 |
+
- tell_a_joke:
|
| 33 |
+
val_split: 0.05
|
| 34 |
+
max_val_set: 250
|
| 35 |
+
- webgpt:
|
| 36 |
+
val_split: 0.05
|
| 37 |
+
max_val_set: 250
|
| 38 |
+
- gpt4all:
|
| 39 |
+
val_split: 0.01
|
| 40 |
+
max_val_set: 1000
|
| 41 |
+
- alpaca_gpt4:
|
| 42 |
+
val_split: 0.025
|
| 43 |
+
max_val_set: 250
|
| 44 |
+
- code_alpaca:
|
| 45 |
+
val_split: 0.05
|
| 46 |
+
max_val_set: 250
|
| 47 |
+
- vicuna:
|
| 48 |
+
max_val_set: 250
|
| 49 |
+
- oig_file:
|
| 50 |
+
source_url: https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl
|
| 51 |
+
max_count: 10000
|
| 52 |
+
min_length: 250
|
| 53 |
+
val_split: 0.05
|
| 54 |
+
max_val_set: 250
|
| 55 |
+
- minimath:
|
| 56 |
+
val_split: 0.05
|
| 57 |
+
- humaneval_mbpp_codegen_qa:
|
| 58 |
+
val_split: 0.05
|
| 59 |
+
- humaneval_mbpp_testgen_qa:
|
| 60 |
+
val_split: 0.05
|
| 61 |
+
- grade_school_math_instructions:
|
| 62 |
+
val_split: 0.05
|
| 63 |
+
- recipes:
|
| 64 |
+
val_split: 0.05
|
| 65 |
+
- cmu_wiki_qa:
|
| 66 |
+
val_split: 0.05
|
| 67 |
+
- oa_wiki_qa_bart_10000row:
|
| 68 |
+
val_split: 0.05
|
| 69 |
+
max_val_set: 250
|
| 70 |
+
- prosocial_dialogue:
|
| 71 |
+
fraction: 0.1
|
| 72 |
+
max_val_set: 250
|
| 73 |
+
- explain_prosocial:
|
| 74 |
+
fraction: 0.075
|
| 75 |
+
max_val_set: 250
|
| 76 |
+
- soda:
|
| 77 |
+
fraction: 0.25
|
| 78 |
+
max_val_set: 1000
|
| 79 |
+
- oa_leet10k:
|
| 80 |
+
val_split: 0.05
|
| 81 |
+
max_val_set: 250
|
| 82 |
+
- dolly15k:
|
| 83 |
+
val_split: 0.05
|
| 84 |
+
max_val_set: 300
|
| 85 |
+
```
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
Pythia:
|
| 89 |
+
```
|
| 90 |
+
pythia-12b-pretrain:
|
| 91 |
+
dtype: fp16
|
| 92 |
+
log_dir: "pythia_log_12b"
|
| 93 |
+
learning_rate: 6e-6
|
| 94 |
+
model_name: EleutherAI/pythia-12b-deduped
|
| 95 |
+
output_dir: pythia_model_12b
|
| 96 |
+
weight_decay: 0.0
|
| 97 |
+
max_length: 2048
|
| 98 |
+
warmup_steps: 100
|
| 99 |
+
gradient_checkpointing: true
|
| 100 |
+
gradient_accumulation_steps: 4
|
| 101 |
+
per_device_train_batch_size: 4
|
| 102 |
+
per_device_eval_batch_size: 4
|
| 103 |
+
eval_steps: 251
|
| 104 |
+
save_steps: 500
|
| 105 |
+
num_train_epochs: 1
|
| 106 |
+
save_total_limit: 2
|
| 107 |
+
deepspeed_config: configs/zero_config_pretrain.json
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
Command used: `deepspeed trainer_sft.py --show_dataset_stats --configs defaults pythia-12b-pretrain pretrain --cache_dir .cache/ --output_dir .saved/pythia-12b-super-pretrain2 --deepspeed`
|