Add inference results README
Browse files- inference/README.md +138 -0
inference/README.md
ADDED
|
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# AGILLM-3 Inference Results
|
| 2 |
+
|
| 3 |
+
## Checkpoint: `pretrain_step10176997.pt`
|
| 4 |
+
- **Step**: 10,176,997
|
| 5 |
+
- **Tokens seen**: 8,367,019,023
|
| 6 |
+
- **Model params**: 698.39M (698,389,088)
|
| 7 |
+
- **Architecture**: AGILLM-3 (d=1024, layers=24, heads=16, rank=128, ALiBi, ReLU FFN 4x)
|
| 8 |
+
- **Tokenizer**: deepseek-ai/DeepSeek-V3.2
|
| 9 |
+
- **Device**: NVIDIA RTX 3090 24GB
|
| 10 |
+
- **Date**: 2026-02-28T22:31:44Z
|
| 11 |
+
- **Mode**: AR (autoregressive)
|
| 12 |
+
- **Temperature**: 0.7
|
| 13 |
+
- **Max new tokens**: 300
|
| 14 |
+
|
| 15 |
+
## Samples
|
| 16 |
+
|
| 17 |
+
### Prompt 1: `The meaning of life is`
|
| 18 |
+
- Tokens: 300 | Speed: 13.7 tok/s | Time: 21.87s
|
| 19 |
+
|
| 20 |
+
```
|
| 21 |
+
a given, but there are already reports of people being killed in the area.
|
| 22 |
+
|
| 23 |
+
One woman living with asthma has said it is "not legally permitted to go to hospital".
|
| 24 |
+
|
| 25 |
+
Paul Robeson was a superstar of the stage and screen, a talented football player and a music hitmaker. Then came a dramatic fall from grace.
|
| 26 |
+
|
| 27 |
+
In a statement, the Spanish singer says the allegations are "absolutely false and deeply sadden me".
|
| 28 |
+
|
| 29 |
+
Labour's Anna Perrett takes on why she object Podium girls to charity.
|
| 30 |
+
|
| 31 |
+
Police say they hav
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
---
|
| 35 |
+
|
| 36 |
+
### Prompt 2: `In a shocking discovery, scientists announced`
|
| 37 |
+
- Tokens: 300 | Speed: 13.9 tok/s | Time: 21.63s
|
| 38 |
+
|
| 39 |
+
```
|
| 40 |
+
the construction of a new building on West Africa's north-west border.
|
| 41 |
+
|
| 42 |
+
The Oscar-winning actor sat down with Stafford after he was warned that 20 or 30 years of fighting for seizing white supremacy in South Sudan.
|
| 43 |
+
|
| 44 |
+
A BBC programme shows councillors there are only a few minutes to stop the return of weapons - but they have been fans in Turkey's capital cities.
|
| 45 |
+
|
| 46 |
+
Raphael Dean is sentenced to count income for ultrasounds as part of ongoing operations near Oban and at St Luke's School Glasgow city
|
| 47 |
+
```
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
### Prompt 3: `def quicksort(arr):`
|
| 52 |
+
- Tokens: 300 | Speed: 14.0 tok/s | Time: 21.42s
|
| 53 |
+
|
| 54 |
+
```
|
| 55 |
+
Se crea el maquillaje de la escena urbana. En algunos momentos, cuando se habla del coronavirus, es muy probable que la gente conozca si se preste a todas las leyes de la cabeza, especialmente en las zonas rurales.
|
| 56 |
+
|
| 57 |
+
Soy redactora de LIFE en El HuffPost Espa帽a, esa secci贸n siempre necesaria en la que mostramos otro lado de la vida m谩s amable y los temas que quiz谩 no tienen cabida y comentarios sobre ella.
|
| 58 |
+
|
| 59 |
+
Escribo sobre todo tipo de temas de actualidad y de temas que te interesan por su magnitud
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
---
|
| 63 |
+
|
| 64 |
+
### Prompt 4: `The Great Wall of China was built`
|
| 65 |
+
- Tokens: 300 | Speed: 13.8 tok/s | Time: 21.74s
|
| 66 |
+
|
| 67 |
+
```
|
| 68 |
+
by the late Tiananmens, and each day its AI interactive assistant built to simulate human reasoning and optimizing them through simple-sounding navigation.
|
| 69 |
+
|
| 70 |
+
"At every stage we celebrate this as well, and for the next generation of AI agents, that will be used in complex tech and manufacturing tasks, from city actors to industrial clusters such as art craftspeople or even influencers providing social support."
|
| 71 |
+
|
| 72 |
+
Former New Zealand PM Samirphos Zardagala , who is also an Uber investor, has been on
|
| 73 |
+
```
|
| 74 |
+
|
| 75 |
+
---
|
| 76 |
+
|
| 77 |
+
### Prompt 5: `Dear Mr. President, I am writing to`
|
| 78 |
+
- Tokens: 300 | Speed: 13.9 tok/s | Time: 21.61s
|
| 79 |
+
|
| 80 |
+
```
|
| 81 |
+
you the first line of that. We have a long history of being present but also a clear and robust woman.
|
| 82 |
+
|
| 83 |
+
"Here is really great fun. Everyone who's listening and even looking forward to our new show," he said in an Instagram post.
|
| 84 |
+
|
| 85 |
+
"We love it all. The good news is that we are doing it when you see your family."
|
| 86 |
+
|
| 87 |
+
The Aida Novochen Foundation says it has been important for people to remember their stress levels.
|
| 88 |
+
|
| 89 |
+
Researchers say more than 350 customers can reverse the effects of direct payments on
|
| 90 |
+
```
|
| 91 |
+
|
| 92 |
+
---
|
| 93 |
+
|
| 94 |
+
### Prompt 6: `Breaking news: astronomers have detected`
|
| 95 |
+
- Tokens: 300 | Speed: 13.9 tok/s | Time: 21.56s
|
| 96 |
+
|
| 97 |
+
```
|
| 98 |
+
a mass number of cases of infectious diseases in the US.
|
| 99 |
+
|
| 100 |
+
The woman, who was formerly infected for several months, told BBC Radio's Good Morning Scotland programme she had been pregnant with her child.
|
| 101 |
+
|
| 102 |
+
Some say they will be "overwhelmed" and not yet trying to get an appointment.El presidente de Estados Unidos, Donald Trump , ha firmado este lunes un decreto para la venta ilegal y ordenada de fara adquisici贸n de Estado Isl谩mico (ISIS por sus siglas en ingl茅s) por el que el republicano no ha fir
|
| 103 |
+
```
|
| 104 |
+
|
| 105 |
+
---
|
| 106 |
+
|
| 107 |
+
### Prompt 7: `Recipe for chocolate cake: Ingredients:`
|
| 108 |
+
- Tokens: 300 | Speed: 13.9 tok/s | Time: 21.56s
|
| 109 |
+
|
| 110 |
+
```
|
| 111 |
+
becomes the healthiest person, including unhealthy and sweet. All of these products will have an emotional impact on your health.
|
| 112 |
+
|
| 113 |
+
Your family's favorite beverage will contain a variety of foods, especially sugar. When you鈥檙e under their control, you're at increased risk for low blood sugar levels , which can cause reactions to unwanted side effects or overdose.
|
| 114 |
+
|
| 115 |
+
There are plenty of good reasons to eat too much orange juice. If you decide to drink orange juice for weight-plus instead of frozen
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
---
|
| 119 |
+
|
| 120 |
+
### Prompt 8: `Once upon a time, a young dragon`
|
| 121 |
+
- Tokens: 300 | Speed: 13.8 tok/s | Time: 21.74s
|
| 122 |
+
|
| 123 |
+
```
|
| 124 |
+
was allowed into her house. She and her husband were separated at the age of 12.
|
| 125 |
+
|
| 126 |
+
As soon as she arrived with her baby, she left behind multiple rooms, and by that point they had saved their lives. They hadn鈥檛 seen anything.
|
| 127 |
+
|
| 128 |
+
The United States saw cases followed in other countries. In Italy, the Red Cross finally opened to authorities in Germany over Christmas Eve, but following an attack by police forces for citizens aged between two and four days, according to reports from Austria and Finland
|
| 129 |
+
```
|
| 130 |
+
|
| 131 |
+
---
|
| 132 |
+
|
| 133 |
+
## Notes
|
| 134 |
+
- This is a **pure pre-training checkpoint** (no SFT/RLHF)
|
| 135 |
+
- The model generates fluent text but does not follow prompts coherently
|
| 136 |
+
- Language mixing (English/Spanish) reflects multilingual crawl data
|
| 137 |
+
- Strong bias toward news/editorial text from training distribution
|
| 138 |
+
- Code prompts produce natural language, not code
|