GeocentricAI/ Geocentric-Arc-1:latest

4 1 week ago

Arc is a small, experimental English model from Geocentric (120.8M parameters, 1,024-token context, 483 MB) for co-writing: brainstorming phrasing and drafting short pieces. Runs offline on almost any machine. It can invent facts. Apache 2.0. 

ollama run GeocentricAI/Geocentric-Arc-1

Details

1 week ago

de6d318b1f4d ยท 485MB

json
{ "architectures": [ "LlamaForCausalLM" ], "model_type": "llama", "vocab_size": 32000, "hidden_size"
json
{ "bos_token_id": 2, "eos_token_id": 7, "pad_token_id": 0 }
json
{ "bos_token": "<bos>", "eos_token": "<|eot|>", "pad_token": "<pad>", "unk_token": "<unk>" }
json
{ "version": "1.0", "truncation": null, "padding": null, "added_tokens": [ { "id": 0, "content": "<p
json
{ "tokenizer_class": "PreTrainedTokenizerFast", "model_max_length": 1024, "chat_template": "{% for m
template
{{- if .System }}<|system|> {{ .System }}<|eot|> {{ end }}<|user|> {{ .Prompt }}<|eot|> <|assistant|
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "num_ctx": 1024, "stop": [ "<|eot|>" ], "temperature": 0.8 }
101 tensors

Readme

Geocentric Arc

Arc is a small, experimental English language model from Geocentric, trained from scratch. It is built for co-writing: exploring phrasing, drafting loose story material and finding fragments to develop in your own voice.

It has 120.8 million parameters, so it runs fast on almost any machine, offline, with nothing sent to anyone.

ollama run geocentricai/arc

At a glance
Parameters 120,787,072
Context 1,024 tokens
Size about 483 MB
Language English
License Apache 2.0
Speed about 190 tokens/second on an Apple M4

Good for

  • Brainstorming phrasing and story ideas
  • Short drafts you plan to revise
  • Experimenting with and studying a small model

Not good for

Arc is experimental. It invents facts, gets arithmetic wrong, and often ignores requested length, format or premise. It can fall into repetition on open-ended prompts. Do not rely on it for facts, calculations, code you will run unreviewed, or anything safety-critical.

In our diagnostic suite of 24 prompts it passed 1, partially passed 4 and failed 19. Every response, including the failures, is published unedited on the model card.

Prompt format

Arc was trained on a simple chat format, which the Ollama build applies for you:

<|user|>
Your message<|eot|>
<|assistant|>

More