tahauzumcu
commited on
Commit
•
a042dde
1
Parent(s):
be44e8d
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,101 @@
|
|
1 |
-
---
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
library_name: peft
|
3 |
+
base_model: unsloth/llama-3-8b-bnb-4bit
|
4 |
+
---
|
5 |
+
|
6 |
+
# VeriUS LLM 8b v0.2
|
7 |
+
|
8 |
+
VeriUS LLM is a generative model that is fine-tuned on Llama-3-8B (Unsloth).
|
9 |
+
|
10 |
+
|
11 |
+
## Model Details
|
12 |
+
Base Model: unsloth/llama-3-8b-bnb-4bit
|
13 |
+
|
14 |
+
Training Dataset: A combined dataset of alpaca, dolly and bactrainx which is translated to turkish.
|
15 |
+
|
16 |
+
Training Method: Fine-tuned with Unsloth, which uses QLoRA. Used ORPO
|
17 |
+
|
18 |
+
#TrainingArguments
|
19 |
+
PER_DEVICE_BATCH_SIZE: 2
|
20 |
+
GRADIENT_ACCUMULATION_STEPS: 4
|
21 |
+
WARMUP_RATIO: 0.03
|
22 |
+
NUM_EPOCHS: 2
|
23 |
+
LR: 0.000008
|
24 |
+
OPTIM: "adamw_8bit"
|
25 |
+
WEIGHT_DECAY: 0.01
|
26 |
+
LR_SCHEDULER_TYPE: "linear"
|
27 |
+
BETA: 0.1
|
28 |
+
|
29 |
+
#PEFT Arguments
|
30 |
+
RANK: 128
|
31 |
+
TARGET_MODULES:
|
32 |
+
- "q_proj"
|
33 |
+
- "k_proj"
|
34 |
+
- "v_proj"
|
35 |
+
- "o_proj"
|
36 |
+
- "gate_proj"
|
37 |
+
- "up_proj"
|
38 |
+
- "down_proj"
|
39 |
+
|
40 |
+
LORA_ALPHA: 256
|
41 |
+
LORA_DROPOUT: 0
|
42 |
+
BIAS: "none"
|
43 |
+
GRADIENT_CHECKPOINT: 'unsloth'
|
44 |
+
USE_RSLORA: false
|
45 |
+
|
46 |
+
## Usage
|
47 |
+
This model is trained used Unsloth and uses it for fast inference. For Unsloth installation please refer to: https://github.com/unslothai/unsloth
|
48 |
+
|
49 |
+
This model can also be loaded with AutoModelForCausalLM
|
50 |
+
|
51 |
+
How to load with unsloth:
|
52 |
+
```commandline
|
53 |
+
from unsloth import FastLanguageModel
|
54 |
+
|
55 |
+
max_seq_len = 2048
|
56 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
57 |
+
model_name="VeriUs/VeriUS-LLM-8b-v0.2",
|
58 |
+
max_seq_length=max_seq_len,
|
59 |
+
dtype=None,
|
60 |
+
load_in_4bit=True
|
61 |
+
)
|
62 |
+
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
|
63 |
+
|
64 |
+
prompt_tempate = """Aşağıda, görevini açıklayan bir talimat ve daha fazla bağlam sağlayan bir girdi verilmiştir. İsteği uygun bir şekilde tamamlayan bir yanıt yaz.
|
65 |
+
|
66 |
+
### Talimat:
|
67 |
+
{}
|
68 |
+
|
69 |
+
### Girdi:
|
70 |
+
{}
|
71 |
+
|
72 |
+
### Yanıt:
|
73 |
+
"""
|
74 |
+
|
75 |
+
|
76 |
+
def generate_output(instruction, user_input):
|
77 |
+
input_ids = tokenizer(
|
78 |
+
[
|
79 |
+
prompt_tempate.format(instruction, user_input)
|
80 |
+
], return_tensors="pt").to("cuda")
|
81 |
+
|
82 |
+
outputs = model.generate(**input_ids, max_length=max_seq_len, do_sample=True)
|
83 |
+
|
84 |
+
# removes prompt, comment this out if you want to see it.
|
85 |
+
outputs = [output[len(input_ids[i].ids):] for i, output in enumerate(outputs)]
|
86 |
+
|
87 |
+
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
88 |
+
|
89 |
+
|
90 |
+
response = generate_output("Türkiye'nin en kalabalık şehri hangisidir?", "")
|
91 |
+
print(response)
|
92 |
+
```
|
93 |
+
|
94 |
+
## Bias, Risks, and Limitations
|
95 |
+
|
96 |
+
Limitations and Known Biases
|
97 |
+
Primary Function and Application: VeriUS LLM, an autoregressive language model, is primarily designed to predict the next token in a text string. While often used for various applications, it is important to note that it has not undergone extensive real-world application testing. Its effectiveness and reliability across diverse scenarios remain largely unverified.
|
98 |
+
|
99 |
+
Language Comprehension and Generation: The base model is primarily trained in standard English. Even though it fine-tuned with and Turkish dataset, its performance in understanding and generating slang, informal language, or other languages may be limited, leading to potential errors or misinterpretations.
|
100 |
+
|
101 |
+
Generation of False Information: Users should be aware that VeriUS LLM may produce inaccurate or misleading information. Outputs should be considered as starting points or suggestions rather than definitive answers.
|