lilbase (297m, trained from scratch) finetuned for chat on smol-smoltalk.

ollama run navthings/lilchat

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

from ollama import chat

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

import ollama from 'ollama'

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

Details

Updated 3 weeks ago

3 weeks ago

f8019b60bb41 · 379MB

model

archllama

parameters297M

quantizationQ8_0

379MB

template

{{- if .System }}<<SYS>> {{ .System }} <</SYS>> {{ end }} {{- range .Messages }} {{- if eq .Role "us

213B

params

{ "num_ctx": 1024, "num_predict": 400, "repeat_penalty": 1.1, "stop": [ "[IN

103B

Readme

lilchat

lilbase (297m params, trained from scratch) finetuned into a chat model on smol-smoltalk.

it follows the chat format, answers in full sentences, stops when it’s done, and does lists and code blocks. it does not know things. expect confident wrong answers on facts and maths, and weak memory across turns.

ollama run navthings/lilchat

tag	size	notes
`latest` / `q8_0`	379mb	same quality as f16
`q4_k_m`	273mb	smallest
`f16`	594mb	unquantized

finetune

full sft in mlx on a macbook air m5, 1,136 steps of 32k tokens (~37m tokens, about 12% of smol-smoltalk), loss only on assistant turns. llama-2 [INST] format, system prompts via <<SYS>>. held-out loss 1.337 vs 2.230 for lilbase on the same data.

defaults

temperature 0.7, top-k 40, repeat penalty 1.1, 400 tokens, 1024 context. greedy decoding loops, keep some temperature.

code

https://github.com/navthings/lilbase (mac branch: mac/sft.py, mac/chat.py)