Neko-Institute-of-Science
commited on
Commit
·
281ca8e
1
Parent(s):
f262779
Add more info
Browse files
README.md
CHANGED
@@ -2,4 +2,45 @@
|
|
2 |
datasets:
|
3 |
- CheshireAI/guanaco-unchained
|
4 |
---
|
5 |
-
Let's see how this goes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
datasets:
|
3 |
- CheshireAI/guanaco-unchained
|
4 |
---
|
5 |
+
Let's see how this goes.
|
6 |
+
|
7 |
+
Training in 8 bit and at full context. Is 8bit even a qlora?
|
8 |
+
|
9 |
+
```
|
10 |
+
python qlora.py \
|
11 |
+
--model_name_or_path /UI/text-generation-webui/models/llama-30b \
|
12 |
+
--output_dir ./output/guanaco-33b \
|
13 |
+
--logging_steps 1 \
|
14 |
+
--save_strategy steps \
|
15 |
+
--data_seed 42 \
|
16 |
+
--save_steps 69 \
|
17 |
+
--save_total_limit 999 \
|
18 |
+
--per_device_eval_batch_size 1 \
|
19 |
+
--dataloader_num_workers 3 \
|
20 |
+
--group_by_length \
|
21 |
+
--logging_strategy steps \
|
22 |
+
--remove_unused_columns False \
|
23 |
+
--do_train \
|
24 |
+
--do_eval false \
|
25 |
+
--do_mmlu_eval false \
|
26 |
+
--lora_r 64 \
|
27 |
+
--lora_alpha 16 \
|
28 |
+
--lora_modules all \
|
29 |
+
--bf16 \
|
30 |
+
--bits 8 \
|
31 |
+
--warmup_ratio 0.03 \
|
32 |
+
--lr_scheduler_type constant \
|
33 |
+
--gradient_checkpointing \
|
34 |
+
--gradient_accumulation_steps 32 \
|
35 |
+
--dataset oasst1 \
|
36 |
+
--source_max_len 2048 \
|
37 |
+
--target_max_len 2048 \
|
38 |
+
--per_device_train_batch_size 1 \
|
39 |
+
--num_train_epochs 3 \
|
40 |
+
--learning_rate 0.0001 \
|
41 |
+
--adam_beta2 0.999 \
|
42 |
+
--max_grad_norm 0.3 \
|
43 |
+
--lora_dropout 0.05 \
|
44 |
+
--weight_decay 0.0 \
|
45 |
+
--seed 0
|
46 |
+
```
|