Microllama_Char_500k_step

This is a character (english a-z 0-9 and so on) trained model following Andrej karpathy's llama.c project https://github.com/karpathy/llama2.c on both TinyStories and my own internal similar dataset I made. the last 150k is from a subset of cosmopedia I extracted for younger people.

Trained for 49,152,000,000 tokens

for it to see/output Uppercase letters this model uses a Shift-Key modifier before the letter to become uppercase, and has never been trained on actual uppercase letters.

This modifier is ↨ and here are the functions I use to convert from straight text to the modified format and back.

def add_caseifer(text):
    # Using list comprehension for more efficient concatenation
    return ''.join(['↨' + char.lower() if char.isupper() else char for char in text
    
def remove_caseifer(text):
    new_text = ""
    i = 0
    while i < len(text):
        if text[i] == "↨":
            if i+1 < len(text):
                new_text += text[i+1].upper()
                i += 1
            else:
                pass  # skip this index
        else:
            new_text += text[i]
        i += 1
    return new_text

As such for test strings to use in chat try using somthing like:

↨hello, my name is ↨clara and ↨i like

Run history: iter ▁▁▁▂▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███ loss/train █▅▄▅▃▅▄▃▄▄▄▄▃▃▄▄▂▁▂▂▃▃▂▃▃▃▂▃▂▃▂▂▂▂▃▁▂▂▁▂ loss/val █▇▆▅▅▄▄▄▄▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁ lr ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ mfu ▅▅▅▄▅▅▅▅▅▅▄▅▅▅▁▅▅▅▄▅▅▅▅█▅▅▅▅▅▅▅▅▅▅▅▅█▅▅█ step_time ██████▇████████▇████▇██▂▇██▇█████▇▇▇▁▇█▁ tokens ▁▁▁▂▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███

Run summary: iter 500000 loss/train 0.48935 loss/val 0.45042 lr 1e-05 mfu 9.31042 step_time 63441.47873 tokens 49152000000