Update README.md
Browse files
README.md
CHANGED
@@ -1,67 +1,72 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
|
5 |
-
Modelo entrenado con DPO
|
6 |
|
7 |
-
Merge de dos modelos
|
8 |
|
9 |
-
|
10 |
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
-
|
22 |
-
|
23 |
-
model_name,
|
24 |
-
torch_dtype=torch.float16,
|
25 |
-
load_in_4bit=True
|
26 |
-
)
|
27 |
-
model.config.use_cache = False
|
28 |
|
29 |
-
|
30 |
-
ref_model = AutoModelForCausalLM.from_pretrained(
|
31 |
-
model_name,
|
32 |
-
torch_dtype=torch.float16,
|
33 |
-
load_in_4bit=True
|
34 |
-
)
|
35 |
|
36 |
-
# Training arguments
|
37 |
-
training_args = TrainingArguments(
|
38 |
-
per_device_train_batch_size=4,
|
39 |
-
gradient_accumulation_steps=4,
|
40 |
-
gradient_checkpointing=True,
|
41 |
-
learning_rate=5e-5,
|
42 |
-
lr_scheduler_type="cosine",
|
43 |
-
max_steps=200,
|
44 |
-
save_strategy="no",
|
45 |
-
logging_steps=1,
|
46 |
-
output_dir=new_model,
|
47 |
-
optim="paged_adamw_32bit",
|
48 |
-
warmup_steps=100,
|
49 |
-
bf16=True,
|
50 |
-
report_to="wandb",
|
51 |
-
)
|
52 |
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
)
|
65 |
|
66 |
-
|
67 |
-
|
|
|
|
1 |
---
|
2 |
+
tags:
|
3 |
+
- merge
|
4 |
+
- mergekit
|
5 |
+
- lazymergekit
|
6 |
+
- mlabonne/OmniBeagle-7B
|
7 |
+
- flemmingmiguel/MBX-7B-v3
|
8 |
+
- AiMavenAi/AiMaven-Prometheus
|
9 |
+
base_model:
|
10 |
+
- mlabonne/OmniBeagle-7B
|
11 |
+
- flemmingmiguel/MBX-7B-v3
|
12 |
+
- AiMavenAi/AiMaven-Prometheus
|
13 |
license: apache-2.0
|
14 |
---
|
15 |
|
|
|
16 |
|
|
|
17 |
|
18 |
+
# NeuTrixOmniBe-DPO
|
19 |
|
20 |
+
NeuTrixOmniBe-DPO is a merge of the following models using [LazyMergekit](https://colab.research.google.com/drive/1obulZ1ROXHjYLn6PPZJwRR6GzgQogxxb?usp=sharing):
|
21 |
+
```yaml
|
22 |
+
MODEL_NAME = "NeuTrixOmniBe-DPO"
|
23 |
+
yaml_config = """
|
24 |
+
slices:
|
25 |
+
- sources:
|
26 |
+
- model: CultriX/NeuralTrix-7B-dpo
|
27 |
+
layer_range: [0, 32]
|
28 |
+
- model: paulml/OmniBeagleSquaredMBX-v3-7B-v2
|
29 |
+
layer_range: [0, 32]
|
30 |
+
merge_method: slerp
|
31 |
+
base_model: CultriX/NeuralTrix-7B-dpo
|
32 |
+
parameters:
|
33 |
+
t:
|
34 |
+
- filter: self_attn
|
35 |
+
value: [0, 0.5, 0.3, 0.7, 1]
|
36 |
+
- filter: mlp
|
37 |
+
value: [1, 0.5, 0.7, 0.3, 0]
|
38 |
+
- value: 0.5
|
39 |
+
dtype: bfloat16
|
40 |
+
"""```
|
41 |
|
42 |
+
It was then trained with DPO using:
|
43 |
+
* Intel/orca_dpo_pairs
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
+
## 🧩 Configuration
|
|
|
|
|
|
|
|
|
|
|
46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
+
|
49 |
+
## 💻 Usage
|
50 |
+
|
51 |
+
```python
|
52 |
+
!pip install -qU transformers accelerate
|
53 |
+
|
54 |
+
from transformers import AutoTokenizer
|
55 |
+
import transformers
|
56 |
+
import torch
|
57 |
+
|
58 |
+
model = "Kukedlc/NeuTrixOmniBe-DPO"
|
59 |
+
messages = [{"role": "user", "content": "What is a large language model?"}]
|
60 |
+
|
61 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
62 |
+
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
63 |
+
pipeline = transformers.pipeline(
|
64 |
+
"text-generation",
|
65 |
+
model=model,
|
66 |
+
torch_dtype=torch.float16,
|
67 |
+
device_map="auto",
|
68 |
)
|
69 |
|
70 |
+
outputs = pipeline(prompt, max_new_tokens=128, do_sample=True, temperature=0.5, top_k=50, top_p=0.95)
|
71 |
+
print(outputs[0]["generated_text"])
|
72 |
+
```
|