PEFT
flan
opt
File size: 2,472 Bytes
17cd953
 
 
 
 
748da8d
 
 
 
 
17cd953
 
c93d072
17cd953
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
---
tags:
- flan
- opt
- peft
datasets:
- SirNeural/flan_v2
metrics:
- perplexity
base_model: facebook/opt-2.7b
---

## FLAN-OPT-2.7b-LoRA

OPT was first introduced in [Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) and first released in [metaseq's repository](https://github.com/facebookresearch/metaseq) on May 3rd 2022 by Meta AI.

This model is [facebook/opt-2.7b](https://hf.co/facebook/opt-2.7b) finetuned with low-rank adapters (https://arxiv.org/abs/2106.09685) on the FLAN datasets (https://arxiv.org/pdf/2210.11416.pdf).

Low-rank adapters (r=16) finetuned over 1.1m new tokens of a FLAN task mixture, with the start of each example cut off if it was too large to fit within a 256 token context.

The model reaches a train ppl of 5.09 and an eval ppl of 4.36.

### Inference Example (Chain-of-Thought prompt):

```python
# %pip install -qq transformers git+https://github.com/huggingface/peft accelerate bitsandbytes
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
peft_model_id = "crumb/FLAN-OPT-2.7b-LoRA"

config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_8bit=True, low_cpu_mem_usage=True, device_map='auto')
model = PeftModel.from_pretrained(model, peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

import torch
prompt = """
Q: Answer the following yes/no question by reasoning step-by-step. Could a dandelion suffer from hepatitis?
A: Hepatitis only affects organisms with livers. Dandelions don’t have a liver. The answer is no.

Q: Answer the following yes/no question by reasoning step-by-step. Can you write a whole Haiku in a single tweet?
A: A haiku is a japanese three-line poem. That is short enough to fit in 280 characters. The answer is yes.

Q: Answer the following yes/no question by reasoning step-by-step. Can you reach space with a Cessna?
A: 
""".strip()
inputs = tokenizer([prompt], return_tensors='pt')

with torch.autocast("cuda", dtype=torch.float16):
    outputs = model.generate(
        input_ids=inputs.input_ids.cuda(),
        attention_mask=inputs.attention_mask.cuda(),
        max_new_tokens=32,
        top_k=4,
        penalty_alpha=0.6
    )
print("\n".join(tokenizer.decode(outputs[0]).split("\n")[:prompt.count("\n")+1]))
# Cessna is a single-engine aircraft. Cessna cannot reach space. The answer is no.
```