Update README.md
Browse files
README.md
CHANGED
|
@@ -16,6 +16,48 @@ language:
|
|
| 16 |
- **Developed by:** alibidaran
|
| 17 |
- **License:** apache-2.0
|
| 18 |
- **Finetuned from model :** unsloth/meta-llama-3.1-8b-instruct-unsloth-bnb-4bit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
| 21 |
|
|
|
|
| 16 |
- **Developed by:** alibidaran
|
| 17 |
- **License:** apache-2.0
|
| 18 |
- **Finetuned from model :** unsloth/meta-llama-3.1-8b-instruct-unsloth-bnb-4bit
|
| 19 |
+
- **Finedtuned with SFT Algorithm**
|
| 20 |
+
## Direct Usages:
|
| 21 |
+
from transformers import TextStreamer
|
| 22 |
+
from unsloth import FastLanguageModel
|
| 23 |
+
import torch
|
| 24 |
+
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
|
| 25 |
+
dtype = 'Bfloat16' # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
| 26 |
+
load_in_4bit = True
|
| 27 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
| 28 |
+
model_name ="alibidaran/LLAMA3-instructive_reasoning",
|
| 29 |
+
max_seq_length = max_seq_length,
|
| 30 |
+
#dtype = dtype,
|
| 31 |
+
load_in_4bit = load_in_4bit,
|
| 32 |
+
#fast_inference = True, # Enable vLLM fast inference
|
| 33 |
+
max_lora_rank = 128,
|
| 34 |
+
gpu_memory_utilization = 0.6, # Reduce if out of memory
|
| 35 |
+
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
|
| 36 |
+
)
|
| 37 |
+
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
|
| 38 |
+
system_prompt="""
|
| 39 |
+
You are a reasonable expert who thinks and answer the users question.
|
| 40 |
+
Before respond first think and create a chain of thoughts in your mind.
|
| 41 |
+
Then respond to the client.
|
| 42 |
+
Your chain of thought and reflection must be in <thinking>..</thinking> format and your respond
|
| 43 |
+
should be in the <output>..</output> format.
|
| 44 |
+
"""
|
| 45 |
+
|
| 46 |
+
messages = [
|
| 47 |
+
{'role':'system','content':system_prompt},
|
| 48 |
+
{"role": "user", "content":'How many r has the word of strawberry?' },
|
| 49 |
+
|
| 50 |
+
]
|
| 51 |
+
inputs = tokenizer.apply_chat_template(
|
| 52 |
+
messages,
|
| 53 |
+
tokenize = True,
|
| 54 |
+
add_generation_prompt = True, # Must add for generation
|
| 55 |
+
return_tensors = "pt",
|
| 56 |
+
).to("cuda")
|
| 57 |
+
|
| 58 |
+
text_streamer = TextStreamer(tokenizer, skip_prompt = True)
|
| 59 |
+
_ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens =2048,
|
| 60 |
+
use_cache = True, temperature = 0.7, min_p = 0.9)
|
| 61 |
|
| 62 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
| 63 |
|