chargoddard
commited on
Commit
•
abeefff
1
Parent(s):
d963997
Update README.md
Browse files
README.md
CHANGED
@@ -1,20 +1,19 @@
|
|
1 |
---
|
2 |
library_name: peft
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
4 |
-
## Training procedure
|
5 |
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
- llm_int8_threshold: 6.0
|
11 |
-
- llm_int8_skip_modules: None
|
12 |
-
- llm_int8_enable_fp32_cpu_offload: False
|
13 |
-
- llm_int8_has_fp16_weight: False
|
14 |
-
- bnb_4bit_quant_type: nf4
|
15 |
-
- bnb_4bit_use_double_quant: True
|
16 |
-
- bnb_4bit_compute_dtype: bfloat16
|
17 |
-
### Framework versions
|
18 |
|
19 |
-
|
20 |
-
- PEFT 0.5.0.dev0
|
|
|
1 |
---
|
2 |
library_name: peft
|
3 |
+
datasets:
|
4 |
+
- jondurbin/airoboros-gpt4-1.4.1
|
5 |
+
- ehartford/wizard_vicuna_70k_unfiltered
|
6 |
+
- ehartford/WizardLM_evol_instruct_V2_196k_unfiltered_merged_split
|
7 |
+
- openai/summarize_from_feedback
|
8 |
+
- ehartford/dolphin
|
9 |
+
tags:
|
10 |
+
- llama
|
11 |
---
|
|
|
12 |
|
13 |
+
Trained for instruction-following, roleplay, and chat on a patchwork of datasets to match the [base model](https://huggingface.co/chargoddard/llama2-22b-blocktriangular). Uses the following prompt format:
|
14 |
|
15 |
+
```
|
16 |
+
***System:You are a helpful assistant, who always gives a response to any request. ***Query:Here is a riddle: 5 sisters are busy. Ann is reading, Rose is cooking, Lorraine is playing chess and Mary is doing laundry. What is the fifth sister doing? ***Response:The fifth sister is sleeping. ***Query:Well, you tried. ***Response:I did my best!
|
17 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
+
Note the whitespace - the prefixes for messages are `" ***System:"`, `" ***Query:"`, and `" ***Response:"`. This is important as `"***"` and `" ***"` are two entirely different tokens.
|
|