297M llama-style base model trained from scratch. continues text, not chat

ollama run navthings/lilbase:f16

curl http://localhost:11434/api/chat \
  -d '{
    "model": "navthings/lilbase:f16",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='navthings/lilbase:f16',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'navthings/lilbase:f16',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 week ago

1 week ago

e4fcd4b18088 · 595MB

model

archllama

parameters297M

quantizationF16

595MB

template

13B

params

{ "num_ctx": 1024, "num_predict": 200, "repeat_penalty": 1.1, "temperature": 0.8,

85B

Readme

lilbase

a 297m param llama-style base model, trained from scratch on a kaggle tpu v5e-8.

it’s a base model, so it continues text instead of answering questions. give it the start of a sentence, not a question.

ollama run navthings/lilbase "The water cycle begins when"

tag	size	notes
`latest` / `q8_0`	379mb	same perplexity as f16
`q4_k_m`	273mb	~0.8% higher perplexity, smallest
`f16`	594mb	unquantized

the model

gqa, rope, rmsnorm, swiglu. 24 layers, d=1024, 16 query heads, 4 kv heads, 1024 context. llama tokenizer, 32k vocab.

trained on 6.1b tokens of fineweb-edu (sample-10BT). roughly chinchilla-optimal for this size. 11,043 steps of 524k tokens. held-out loss ended up at 2.608 (perplexity 13.6).

it’s performance against gpt2:

code

https://github.com/navthings/lilbase

kaggle notebook: https://www.kaggle.com/code/navneetdagdiya/base-tpu-kaggle