Philipp-Sc
commited on
Commit
•
96aa9b7
1
Parent(s):
209c10c
Update README.md
Browse files
README.md
CHANGED
@@ -15,15 +15,14 @@ base_model: mistralai/Mistral-7B-v0.2
|
|
15 |
(>=checkpoint-v2)
|
16 |
---
|
17 |
|
18 |
-
## Reverse Instruct
|
19 |
|
20 |
-
This
|
21 |
|
22 |
## Response Format
|
23 |
|
24 |
"[INST]\n### System:\n{system}\n### Instruction:\n{instruction}\n[/INST]\n"
|
25 |
|
26 |
-
(without the "")
|
27 |
|
28 |
## Prompt Template
|
29 |
|
@@ -48,6 +47,9 @@ CUDA_VISIBLE_DEVICES=0 WANDB_DISABLED=True python LLaMA-Factory/src/train_bash.p
|
|
48 |
--stage sft \
|
49 |
--model_name_or_path model_name_or_path \
|
50 |
--checkpoint_dir checkpoint_dir \
|
|
|
|
|
|
|
51 |
--do_train \
|
52 |
--dataset default \
|
53 |
--template vanilla \
|
|
|
15 |
(>=checkpoint-v2)
|
16 |
---
|
17 |
|
18 |
+
## Reverse Instruct LoRA Adapter
|
19 |
|
20 |
+
This LoRA Adapter is fine tuned to reverse engineer the original prompt of a given LLM output/response.
|
21 |
|
22 |
## Response Format
|
23 |
|
24 |
"[INST]\n### System:\n{system}\n### Instruction:\n{instruction}\n[/INST]\n"
|
25 |
|
|
|
26 |
|
27 |
## Prompt Template
|
28 |
|
|
|
47 |
--stage sft \
|
48 |
--model_name_or_path model_name_or_path \
|
49 |
--checkpoint_dir checkpoint_dir \
|
50 |
+
--flash_attn \
|
51 |
+
--shift_attn \
|
52 |
+
--neftune_noise_alpha 5 \
|
53 |
--do_train \
|
54 |
--dataset default \
|
55 |
--template vanilla \
|