File size: 2,931 Bytes
ddf3266
8b7ebce
 
 
 
 
 
 
 
ddf3266
7c31a27
 
 
 
 
 
 
 
 
 
 
c8b4d6c
8b7ebce
 
 
 
 
 
 
 
 
9626c37
8b7ebce
 
 
 
 
 
 
 
 
 
9626c37
 
 
 
 
8b7ebce
6edf637
265b5f0
6edf637
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
265b5f0
 
 
 
 
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
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
---
license: cc-by-nc-4.0
language:
- en
tags: 
- text-generation
datasets: 
- stanford_alpaca
pipeline_tag: text-generation
---

<br><br>

<p align="center">
<img src="https://github.com/jina-ai/finetuner/blob/main/docs/_static/finetuner-logo-ani.svg?raw=true" alt="Finetuner logo: Finetuner helps you to create experiments in order to improve embeddings on search tasks. It accompanies you to deliver the last mile of performance-tuning for neural search applications." width="150px">
</p>


<p align="center">
<b>LLM Generation models trained by Jina AI, Finetuner team.</b>
  
This repo contains the full weights (16bit) for Falcon-40b
fit on the [Code Alpaca](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k) dataset.

## Reproduction
This version of the weights was trained with the following hyperparameters:

- Epochs: 2
- Batch size: 128
- Micro batch size: 4
- Learning rate: 3e-4
- Lora _r_: 8
- Lora target modules: query_key_value

You can reproduce using this repository: 

https://github.com/jina-ai/jerboa

Make sure you install requirements and finetune using this command using the following command:

```
python finetune.py \
--base-model tiiuae/falcon-40b --lora-target-modules query_key_value \
--data-path sahil2801/CodeAlpaca-20k --output-dir ./lora-alpaca-code \
--batch-size 128 --micro-batch-size 4 --eval-limit 45 \
--eval-file code_eval.jsonl --wandb-project jerboa --wandb-log-model \
--wandb-watch gradients --num-epochs 2
```

## Inference

```Python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM


TOKENIZER_SOURCE = 'tiiuae/falcon-40b'
BASE_MODEL = 'jinaai/falcon-40b-code-alpaca'
DEVICE = "cuda"

PROMPT = """
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Write a for loop in python

### Input:

### Response:
"""
model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=BASE_MODEL,
    torch_dtype=torch.float16,
    trust_remote_code=True,
    device_map='auto',
)

model.eval()

tokenizer = AutoTokenizer.from_pretrained(
    TOKENIZER_SOURCE,
    trust_remote_code=True,
    padding_side='left',
)
tokenizer.pad_token = tokenizer.eos_token

inputs = tokenizer(PROMPT, return_tensors="pt")
input_ids = inputs["input_ids"].to(DEVICE)
input_attention_mask = inputs["attention_mask"].to(DEVICE)

with torch.no_grad():
    generation_output = model.generate(
        input_ids=input_ids,
        attention_mask=input_attention_mask,
        return_dict_in_generate=True,
        max_new_tokens=32,
        eos_token_id=tokenizer.eos_token_id,
    )
generation_output = generation_output.sequences[0]
output = tokenizer.decode(generation_output, skip_special_tokens=True)

print(output)

```

## Contact

Join our [Discord community](https://discord.jina.ai) and chat with other community members about ideas.