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:
👉 [https://github.com/PursuitOfDataScience/ArgonneAI/tree/Argonne-1.5](https://github.com/PursuitOfDataScience/ArgonneAI/tree/Argonne-1.5)
---
## Model Improvements 🚀
Compared to Argonne-1.0, this model features significant enhancements, including:
- Utilized `torch.compile()` for improved pretraining speed.
- Integrated Flash Attention for an additional **2.6× 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 📚
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 🖥️
- Trained on a single DGX node (8× A100 GPUs, 80GB each).
- **Total GPU Hours:** 1248
- **Total Global Steps:** 80,000
---
## How to Use (Inference) ⚙️
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) |
---
|