instruction-pretrain
commited on
Commit
•
cde5d71
1
Parent(s):
14ebbbf
Update README.md
Browse files
README.md
CHANGED
@@ -29,12 +29,14 @@ We explore supervised multitask pre-training by proposing ***Instruction Pre-Tra
|
|
29 |
- Domain-Specific Models Pre-Trained from Llama3-8B:
|
30 |
- [Finance-Llama3-8B](https://huggingface.co/instruction-pretrain/finance-Llama3-8B)
|
31 |
- [Biomedicine-Llama3-8B](https://huggingface.co/instruction-pretrain/medicine-Llama3-8B)
|
|
|
|
|
32 |
|
33 |
|
34 |
## Domain-Adaptive Continued Pre-Training
|
35 |
Following [AdaptLLM](https://huggingface.co/AdaptLLM/finance-chat), we augment the domain-specific raw corpora with instruction-response pairs generated by our [context-based instruction synthesizer](https://huggingface.co/instruction-pretrain/instruction-synthesizer).
|
36 |
|
37 |
-
|
38 |
```python
|
39 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
40 |
|
@@ -60,6 +62,31 @@ pred = tokenizer.decode(outputs[answer_start:], skip_special_tokens=True)
|
|
60 |
print(pred)
|
61 |
```
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
## Citation
|
64 |
If you find our work helpful, please cite us:
|
65 |
|
|
|
29 |
- Domain-Specific Models Pre-Trained from Llama3-8B:
|
30 |
- [Finance-Llama3-8B](https://huggingface.co/instruction-pretrain/finance-Llama3-8B)
|
31 |
- [Biomedicine-Llama3-8B](https://huggingface.co/instruction-pretrain/medicine-Llama3-8B)
|
32 |
+
- General Instruction-Augmented Corpora: [general-instruction-augmented-corpora](https://huggingface.co/datasets/instruction-pretrain/general-instruction-augmented-corpora)
|
33 |
+
- Domain-Specific Instruction-Augmented Corpora (no finance data to avoid ethical issues): [medicine-instruction-augmented-corpora](https://huggingface.co/datasets/instruction-pretrain/medicine-instruction-augmented-corpora)
|
34 |
|
35 |
|
36 |
## Domain-Adaptive Continued Pre-Training
|
37 |
Following [AdaptLLM](https://huggingface.co/AdaptLLM/finance-chat), we augment the domain-specific raw corpora with instruction-response pairs generated by our [context-based instruction synthesizer](https://huggingface.co/instruction-pretrain/instruction-synthesizer).
|
38 |
|
39 |
+
### 1. To chat with the finance-Llama3-8B model:
|
40 |
```python
|
41 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
42 |
|
|
|
62 |
print(pred)
|
63 |
```
|
64 |
|
65 |
+
### 2. To evaluate our models on the domain-specific tasks
|
66 |
+
1. Set up dependencies
|
67 |
+
```bash
|
68 |
+
git clone https://github.com/microsoft/LMOps
|
69 |
+
cd LMOps/adaptllm
|
70 |
+
pip install -r requirements.txt
|
71 |
+
```
|
72 |
+
|
73 |
+
2. Evaluate
|
74 |
+
```bash
|
75 |
+
DOMAIN='finance'
|
76 |
+
|
77 |
+
# if the model can fit on a single GPU: set MODEL_PARALLEL=False
|
78 |
+
# elif the model is too large to fit on a single GPU: set MODEL_PARALLEL=True
|
79 |
+
MODEL_PARALLEL=False
|
80 |
+
|
81 |
+
# number of GPUs, chosen from [1,2,4,8]
|
82 |
+
N_GPU=1
|
83 |
+
|
84 |
+
# Set as True
|
85 |
+
add_bos_token=True
|
86 |
+
|
87 |
+
bash scripts/inference.sh ${DOMAIN} 'instruction-pretrain/finance-Llama3-8B' ${add_bos_token} ${MODEL_PARALLEL} ${N_GPU}
|
88 |
+
```
|
89 |
+
|
90 |
## Citation
|
91 |
If you find our work helpful, please cite us:
|
92 |
|