Commit
•
fe5cc7c
1
Parent(s):
39cbf8d
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,67 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
|
5 |
+
|
6 |
+
# Model Card for Emu
|
7 |
+
|
8 |
+
|
9 |
+
Some alignments in these domains:
|
10 |
+
- Bitcoin
|
11 |
+
- Nostr
|
12 |
+
- Health
|
13 |
+
- Permaculture
|
14 |
+
- Phytochemicals
|
15 |
+
- Alternative medicine
|
16 |
+
- Herbs
|
17 |
+
- Nutrition
|
18 |
+
|
19 |
+
I am having success with chat template of Llama3: \<\|begin_of_text\|\>\<\|start_header_id\|\> ...
|
20 |
+
You can check the GGUF chat template to see the exact format. But I didn't change it, so Llama3 format continues.
|
21 |
+
GGUF has the necessary eot token to properly stop.
|
22 |
+
|
23 |
+
|
24 |
+
## Model Details
|
25 |
+
|
26 |
+
- **Fine tuned by:** someone
|
27 |
+
- **Finetuned from model:** https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
|
28 |
+
|
29 |
+
|
30 |
+
## Uses
|
31 |
+
|
32 |
+
Ask any question, compared to other models this may know more about those topics above.
|
33 |
+
You can use llama.cpp to chat with it.
|
34 |
+
You can also use llama-cpp-python package to chat with it in a Python script.
|
35 |
+
|
36 |
+
This is how you generate prompt and stops:
|
37 |
+
|
38 |
+
```
|
39 |
+
prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{sys_msg}<|eot_id|>"
|
40 |
+
i = 0
|
41 |
+
while i < len(msgs):
|
42 |
+
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{msgs[i]['content']}<|eot_id|>"
|
43 |
+
prompt += f"<|start_header_id|>assistant<|end_header_id|>\n\n{msgs[i + 1]['content']}<|eot_id|>"
|
44 |
+
i += 2
|
45 |
+
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{q}<|eot_id|>"
|
46 |
+
prompt += "<|start_header_id|>assistant<|end_header_id|>\n\n"
|
47 |
+
stops = ['<|eot_id|>', '<|end_of_text|>', '<|im_end|>', '<|start_header_id|>']
|
48 |
+
```
|
49 |
+
|
50 |
+
|
51 |
+
## Warning
|
52 |
+
|
53 |
+
Users (both direct and downstream) should be aware of the risks, biases and limitations of the model.
|
54 |
+
The trainer, developer or uploader of this model does not assume any liability. Use it at your own risk.
|
55 |
+
|
56 |
+
|
57 |
+
## Training Details
|
58 |
+
|
59 |
+
### Training Data
|
60 |
+
|
61 |
+
Some data I curated from various sources.
|
62 |
+
|
63 |
+
### Training Procedure
|
64 |
+
|
65 |
+
LLaMa-Factory is used to train on 2x3090!
|
66 |
+
|
67 |
+
fsdp_qlora is the technique.
|