File size: 6,245 Bytes
a5d4833
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
04afbe7
a5d4833
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb949fa
a5d4833
 
 
67f9c92
 
 
 
5d0a4f7
67f9c92
 
04afbe7
67f9c92
 
 
04afbe7
 
 
 
67f9c92
04afbe7
67f9c92
04afbe7
67f9c92
04afbe7
 
 
 
 
67f9c92
04afbe7
 
 
 
 
 
 
67f9c92
04afbe7
 
 
 
67f9c92
04afbe7
 
 
 
67f9c92
04afbe7
 
 
67f9c92
 
04afbe7
 
 
 
 
 
 
 
 
 
67f9c92
04afbe7
 
 
 
 
 
 
67f9c92
 
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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
---
license: mit
datasets:
- HuggingFaceFW/fineweb-edu
language:
- en
pipeline_tag: text-classification
---


# Argonne 1.5

Argonne 1.5 is an optimized and significantly improved causal language model compared to its predecessor, Argonne 1.0. It is built on Hugging Face's Transformers library, fully integrated with the `AutoModel` class for ease of use.

The codebase is in:

πŸ‘‰ [https://github.com/PursuitOfDataScience/ArgonneAI/tree/Argonne-1.5](https://github.com/PursuitOfDataScience/ArgonneAI/tree/Argonne-1.5)

---

## Model Improvements πŸš€

Compared to Argonne-1.0, this model features significant enhancements, including:

- Utilized `torch.compile()` for improved pretraining speed.
- Integrated Flash Attention for an additional **2.6Γ— memory efficiency improvement**, enabling a much larger batch size.
- Increased number of layers and attention heads for richer model capacity.
- Better GPU hardware utilization.
- Hugging Face `AutoModel` integration for convenient use.
- Enhanced text generation capabilities.

---

## Training Data πŸ“š

This model was trained on the **same dataset** used for Argonne-1.0 pretraining. Total processed tokens: 15,453,927,424.

---

## Model Details 🧠

| Attribute           | Value       |
|---------------------|-------------|
| **Parameters**      | 356,516,640 |
| **Block Size**      | 2048        |
| **Layers**          | 16          |
| **Attention Heads** | 16          |
| **Embedding Size**  | 1296        |
| **Batch Size**      | 756         |

---

## Training Hardware & Duration πŸ–₯️

- Trained on a single DGX node (8Γ— A100 GPUs, 80GB each).
- **Total GPU Hours:** 1248
- **Total Global Steps:** 80,000


---

## How to Use (Inference) βš™οΈ

Here's how you can quickly start generating text with Argonne-1.5:

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

model_path = "PursuitOfDataScience/Argonne-1.5"

# 1) Load the Argonne-1.5 model (set trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True
)

# 2) Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)

# 3) Inference example
prompt = "The meaning of life is "
inputs = tokenizer(prompt, return_tensors="pt")

# generate text using common HF parameters
outputs = model.generate(**inputs, max_length=150, do_sample=True, top_k=50, top_p=0.95, temperature=0.7)

# Decode and print the generated text
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

<pre>
The meaning of life is tamed in many ways. It is a state of mental and physical development. It is a state of deep emotional strength and confidence, and it is a state of physical and mental balance. In this article, we will explore the meaning of life, the different ways life is defined, and how we can apply this concept to our own lives.
</pre>

# MMLU Evaluation Results

## **Overall Accuracy (0-shot)**
**0.2549 (3579/14042)**

## Breakdown by Subject (Alphabetical Order):
| Subject | Accuracy (Correct/Total) |
|------------------------------|----------------------|
| abstract_algebra             | 0.3200 (32/100)     |
| anatomy                      | 0.3333 (45/135)     |
| astronomy                    | 0.2566 (39/152)     |
| business_ethics              | 0.2300 (23/100)     |
| clinical_knowledge           | 0.2226 (59/265)     |
| college_biology              | 0.3056 (44/144)     |
| college_chemistry            | 0.2100 (21/100)     |
| college_computer_science     | 0.2700 (27/100)     |
| college_medicine             | 0.2543 (44/173)     |
| college_mathematics          | 0.2700 (27/100)     |
| college_physics              | 0.2255 (23/102)     |
| computer_security            | 0.2900 (29/100)     |
| conceptual_physics           | 0.2213 (52/235)     |
| electrical_engineering       | 0.2759 (40/145)     |
| elementary_mathematics       | 0.2963 (112/378)    |
| econometrics                 | 0.2544 (29/114)     |
| formal_logic                 | 0.1508 (19/126)     |
| global_facts                 | 0.3100 (31/100)     |
| high_school_biology          | 0.2613 (81/310)     |
| high_school_chemistry        | 0.3054 (62/203)     |
| high_school_computer_science | 0.3100 (31/100)     |
| high_school_european_history | 0.2970 (49/165)     |
| high_school_geography        | 0.2626 (52/198)     |
| high_school_government_and_politics | 0.2280 (44/193) |
| high_school_macroeconomics   | 0.2051 (80/390)     |
| high_school_mathematics      | 0.2630 (71/270)     |
| high_school_microeconomics   | 0.2059 (49/238)     |
| high_school_physics          | 0.2384 (36/151)     |
| high_school_psychology       | 0.2220 (121/545)    |
| high_school_statistics       | 0.2222 (48/216)     |
| high_school_us_history       | 0.2549 (52/204)     |
| high_school_world_history    | 0.2658 (63/237)     |
| human_aging                  | 0.2377 (53/223)     |
| human_sexuality              | 0.2137 (28/131)     |
| international_law            | 0.3636 (44/121)     |
| jurisprudence                | 0.2315 (25/108)     |
| logical_fallacies            | 0.2945 (48/163)     |
| machine_learning             | 0.2054 (23/112)     |
| management                   | 0.1845 (19/103)     |
| marketing                    | 0.2436 (57/234)     |
| medical_genetics             | 0.2100 (21/100)     |
| miscellaneous                | 0.2439 (191/783)    |
| moral_disputes               | 0.2803 (97/346)     |
| moral_scenarios              | 0.2469 (221/895)    |
| nutrition                    | 0.2353 (72/306)     |
| philosophy                   | 0.3055 (95/311)     |
| prehistory                   | 0.3025 (98/324)     |
| professional_accounting      | 0.2766 (78/282)     |
| professional_law             | 0.2692 (413/1534)   |
| professional_medicine        | 0.1654 (45/272)     |
| professional_psychology      | 0.2827 (173/612)    |
| public_relations             | 0.2182 (24/110)     |
| security_studies             | 0.2449 (60/245)     |
| sociology                    | 0.2388 (48/201)     |
| us_foreign_policy            | 0.2500 (25/100)     |
| virology                     | 0.2048 (34/166)     |
| world_religions              | 0.3041 (52/171)     |

---