Update README.md
Browse files
README.md
CHANGED
@@ -32,4 +32,58 @@ Finetuned for 1 epoch on an A100 through Vast.AI.
|
|
32 |
|
33 |
## Credits
|
34 |
|
35 |
-
Thank you to Axolotl for making finetuning easier. Thank you to Docker for... existing, I guess.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
## Credits
|
34 |
|
35 |
+
Thank you to Axolotl for making finetuning easier. Thank you to Docker for... existing, I guess.
|
36 |
+
|
37 |
+
## YML Configuration
|
38 |
+
|
39 |
+
```
|
40 |
+
base_model: SicariusSicariiStuff/Impish_Nemo_12B
|
41 |
+
|
42 |
+
load_in_8bit: false
|
43 |
+
load_in_4bit: true
|
44 |
+
adapter: qlora
|
45 |
+
|
46 |
+
gradient_accumulation_steps: 8
|
47 |
+
micro_batch_size: 2
|
48 |
+
num_epochs: 1
|
49 |
+
optimizer: adamw_bnb_8bit
|
50 |
+
lr_scheduler: cosine
|
51 |
+
learning_rate: 0.00001
|
52 |
+
|
53 |
+
sequence_len: 4096
|
54 |
+
|
55 |
+
lora_r: 16
|
56 |
+
lora_alpha: 32
|
57 |
+
lora_dropout: 0.05
|
58 |
+
lora_target_linear: true
|
59 |
+
|
60 |
+
bf16: true
|
61 |
+
tf32: false
|
62 |
+
|
63 |
+
gradient_checkpointing: true
|
64 |
+
gradient_checkpointing_kwargs:
|
65 |
+
use_reentrant: false
|
66 |
+
logging_steps: 1
|
67 |
+
flash_attention: true
|
68 |
+
|
69 |
+
loss_watchdog_threshold: 5.0
|
70 |
+
loss_watchdog_patience: 3
|
71 |
+
|
72 |
+
rl: dpo
|
73 |
+
datasets:
|
74 |
+
- path: sam-paech/gutenberg3-generalfiction-scifi-fantasy-romance-adventure-dpo
|
75 |
+
split: train
|
76 |
+
type: chatml.prompt_pairs
|
77 |
+
- path: nbeerbower/gutenberg-moderne-dpo
|
78 |
+
split: train
|
79 |
+
type: chatml.prompt_pairs
|
80 |
+
- path: nbeerbower/gutenberg2-dpo
|
81 |
+
split: train
|
82 |
+
type: chatml.prompt_pairs
|
83 |
+
- path: jondurbin/gutenberg-dpo-v0.1
|
84 |
+
split: train
|
85 |
+
type: chatml.prompt_pairs
|
86 |
+
dataset_prepared_path: last_run_prepared
|
87 |
+
val_set_size: 0.1
|
88 |
+
output_dir: ./outputs/lora-out
|
89 |
+
```
|