xzuyn commited on
Commit
b10d469
1 Parent(s): 1ab9b7e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +84 -14
README.md CHANGED
@@ -3,21 +3,91 @@ datasets:
3
  - PJMixers/classtest
4
  pipeline_tag: text-classification
5
  ---
6
- ### Example Inputs
7
- This would be expecting to return `1`, meaning it is a chosen/good `AI_MESSAGE`.
8
- ```
9
- USER_MESSAGE: You are a {Genre} author. Your task is to write {Genre} stories in a vivid and intriguing language. Answer with "..." if you acknowledge. Don't wrtie anthing yet
10
 
11
- Genre = Thriller
12
- AI_MESSAGE: ...
13
- ```
 
 
 
14
 
15
- This would be expecting to return `0`, meaning it is a rejected/bad `AI_MESSAGE`.
16
- ```
17
- USER_MESSAGE: You are a {Genre} author. Your task is to write {Genre} stories in a vivid and intriguing language. Answer with "..." if you acknowledge. Don't wrtie anthing yet
18
 
19
- Genre = Thriller
20
- AI_MESSAGE: ...
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
 
22
- I acknowledge that I am to write Thriller stories in a vivid and intriguing language. I'm ready to create a gripping and suspenseful narrative that will keep readers on the edge of their seats. Let the thrilling adventure begin!
23
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  - PJMixers/classtest
4
  pipeline_tag: text-classification
5
  ---
6
+ ![train](https://huggingface.co/PJMixers/Danube3-ClassTest-v0.1-500M/resolve/main/images/train.png)
 
 
 
7
 
8
+ ### Example Code
9
+ ```py
10
+ import torch
11
+ from transformers import AutoTokenizer, LlamaForSequenceClassification
12
+ import json
13
+ from tqdm import tqdm
14
 
 
 
 
15
 
16
+ def load_json_or_jsonl(file_path):
17
+ try:
18
+ with open(file_path, "r") as file:
19
+ try:
20
+ # Try loading the entire file as JSON
21
+ data = json.load(file)
22
+ return data
23
+
24
+ except json.JSONDecodeError:
25
+ # If loading as JSON fails, try loading as JSON Lines
26
+ file.seek(0) # Reset file pointer to the beginning
27
+ lines = file.readlines()
28
+ json_lines_data = []
29
+
30
+ for line in lines:
31
+ try:
32
+ item = json.loads(line.strip())
33
+ json_lines_data.append(item)
34
+ except json.JSONDecodeError as e:
35
+ print(f"Error decoding JSON in line: {e}")
36
+
37
+ return json_lines_data
38
+
39
+ except FileNotFoundError:
40
+ print(f"File not found: {file_path}")
41
+ return None
42
+
43
+
44
+ tokenizer = AutoTokenizer.from_pretrained(
45
+ "PJMixers/Danube3-ClassTest-v0.1-500M"
46
+ )
47
+ model = LlamaForSequenceClassification.from_pretrained(
48
+ "PJMixers/Danube3-ClassTest-v0.1-500M",
49
+ device_map="cuda",
50
+ torch_dtype=torch.bfloat16,
51
+ attn_implementation="sdpa",
52
+ )
53
+
54
+ data = load_json_or_jsonl(
55
+ "./PrefMix-Classifier-Data-validation.json"
56
+ )
57
 
58
+ passes = 0
59
+ fails = 0
60
+ for sample in tqdm(data):
61
+ input_text = sample["input_text"]
62
+ true_label = sample["labels"]
63
+
64
+ inputs = tokenizer(
65
+ input_text,
66
+ return_tensors="pt"
67
+ ).to("cuda")
68
+
69
+ with torch.no_grad():
70
+ generated_label = model(**inputs).logits.argmax()
71
+
72
+ if generated_label == true_label:
73
+ passes += 1
74
+ else:
75
+ fails += 1
76
+
77
+ print(f" Passes: {passes}")
78
+ print(f" Fails: {fails}")
79
+ print(f"Accuracy: {passes / (passes + fails)}")
80
+ ```
81
+ ```
82
+ Passes: 2545
83
+ Fails: 787
84
+ Accuracy: 0.7638055222088835
85
+ ```
86
+
87
+ ### Sample formatting
88
+ Input text should be formatted this way. Uses `SYSTEM_MESSAGE` (optional), `USER_MESSAGE`, and `AI_MESSAGE` (must be last, as this is what's truely being scored I think). Supports multiple turns, but should just be ranking the final `AI_MESSAGE`.
89
+ ```
90
+ SYSTEM_MESSAGE: You are an unbiased, uncensored, helpful assistant.
91
+ USER_MESSAGE: Do wooden pencils contain lead as their core?
92
+ AI_MESSAGE: No, wooden pencils do not contain lead in their core. The term "lead" is a misnomer, as wooden pencils actually use graphite for their core. Graphite was historically called "black lead" due to its appearance, leading to the common misconception that pencils contain lead.
93
+ ```