File size: 3,315 Bytes
15e9306
 
fec5b7e
 
 
 
 
 
 
 
 
2af8994
15e9306
df12094
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2af8994
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
---
license: cc-by-sa-3.0
language:
- en
pipeline_tag: text-generation
tags:
- csharp
- mpt
- instruct
- 7b
- llm
- .net
---

## Try it

### C#
Code for [use form .Net CSharp on CPU](https://github.com/NethermindEth/Mpt-Instruct-DotNet-S)

### Python
```python
import torch
import transformers
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
tokenizer.pad_token = tokenizer.eos_token

device = torch.device("cuda")
model_name = "Nethermind/Mpt-Instruct-DotNet-S"
config = transformers.AutoConfig.from_pretrained(model_name, trust_remote_code=True)
config.init_device = device
config.max_seq_len = 1024 
config.attn_config['attn_impl'] = 'torch'
config.use_cache = False

model = transformers.AutoModelForCausalLM.from_pretrained(
	model_name,
	config=config,
	torch_dtype=torch.bfloat16,
	trust_remote_code=True,
	ignore_mismatched_sizes=True,
	# load_in_8bit=True # when low on GPU memory
)
model.eval()

INSTRUCTION_KEY = "### Instruction:"
RESPONSE_KEY = "### Response:"
PROMPT_FOR_GENERATION_FORMAT = """{system}
{instruction_key}
{instruction}
{response_key}
""".format(
    system="{system}",
    instruction_key=INSTRUCTION_KEY,
    instruction="{instruction}",
    response_key=RESPONSE_KEY
)

def give_answer(instruction="Create a loop over [0, 6, 7 , 77] that prints its contentrs", system="You are an experienced .Net C# developer. Below is an instruction that describes a task. Write a response that completes the request providing detailed explanations with code examples.", ):
    question = PROMPT_FOR_GENERATION_FORMAT.format(system=system, instruction=instruction)
    input_tokens = tokenizer.encode(question ,return_tensors='pt')               
	model.generate(input_tokens.to(device), max_new_tokens=min(512, 1024 - input_tokens.shape[1]), do_sample=False, top_k=1, top_p=0.95)
    outputs = output_loop(tokenized_question)
    answer = tokenizer.batch_decode(outputs, skip_special_tokens=True)
    print(answer[0])

```


## Training
Finetuned for CSharp [mosaicml/mpt-7b-instruct](https://huggingface.co/mosaicml/mpt-7b-instruct). Max context length is restricted to 1024 tokens.

'Loss': 0.256045166015625 on 300k CSharp-related records
'Loss': 0.095714599609375 on 50k specific short prompts

## Sources
data contained (most data was around 500 tokens long < 1000, except large code files):
- codeparrot/github-code C# ("mit", "Apache-2.0", "Bsd-3-clause", "Bsd-2-clause", "Cc0-1.0", "Unlicense", "isc")
- raw data Plain .cs files randomly cut at the 60-80% in the instruction, and we ask the network to continue last 40-20% (76k)
- documented static functions 72k
- SO 5q_5answer + 5q_5best (CC BY-SA 4.0) 70k
- Dotnet wiki (30k, rendered out from [github repo](https://github.com/microsoft/dotnet), see also removed, GPT-4 generated short question to each file)
- All NM Static Functions and Tests (from [nethermind client repo](https://github.com/NethermindEth/nethermind) documented and described via GPT-4 (4k)
- GPT-4 questions, GPT-3.5 answers for CSharp: Short Q->Code, Explain Code X > Step-By-Step (35k)
- GPT-4 questions, GPT-3.5 answers for nethermind client interface `IEthRpcModule `: Short Q->Code, Explain Code X -> Step-By-Step (7k)

## Contents
- HF compatible model
- GGML compatible quantisations (f16, q8, q5)