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 runs on your own machine, and nothing you write is sent to us.
Arc is experimental. It is not ready for dependable factual assistance, arithmetic or autonomous work. Its full evaluation, including the responses it gets wrong, is on the model card.
Model details
| Developer | Geocentric |
| Parameters | 120,787,072 |
| Architecture | Decoder-only transformer: RMSNorm, rotary position embeddings, grouped-query attention, SwiGLU feed-forward, tied embeddings |
| Layers x width | 11 x 896 |
| Attention heads | 14 (2 key/value heads) |
| Context length | 1,024 tokens |
| Vocabulary | 32,000 (byte-level BPE) |
| Stage | Supervised fine-tuned (SFT) |
| Modality | Text only |
| Language | English |
| License | Apache License 2.0 |
Evaluation
Scored with PARALLAX 1.0, the evaluation suite we built for our own models. The Parallax Index is 53.5 / 100, and it reproduced on a second machine (Apple M4).
| Probe | Score | Measured |
|---|---|---|
| ZENITH | 22.7 | 2.042 bits/byte on held-out text |
| MERIDIAN | 100.0 | +0.447 nats gained from context |
| SEXTANT | 26.7 | 18/40 correct (chance is 25%) |
| ASTROLABE | 58.3 | 7/12 instructions followed |
| NADIR | 100.0 | 0 of 6 generations looped |
PARALLAX is a small, purpose-built suite and not a standardized benchmark. On Geocentric's published diagnostic prompts Arc passed 1 of 24, partially passed 4 and failed 19. The unedited responses are on the model card.
Files
| File | What it is |
|---|---|
geocentric_sft_best.pt |
The SFT checkpoint (weights and training state needed to load it) |
config.json |
Model configuration |
tokenizer.json |
The 32,000-token tokenizer |
How to use it
The checkpoint is on Ollama and can be ran there by running the following command
ollama run GeocentricAI/Geocentric-Arc-1
Intended use
Co-writing, experimentation, and research on small models and their evaluation.
Limitations
- Arc invents facts when information is missing, and it failed all four arithmetic cases in our diagnostic suite.
- It does not reliably follow requested length, format or premise.
- Its 1,024-token context is short. Long documents will not fit.
- It reflects its training data, including that data's biases. Do not use it as a source of facts, for anything safety-critical, or unsupervised in front of users.
- Generated code can look correct and be broken. Review and test it.
Training data
Arc was trained on public datasets, not on data we bought, and no user conversations were used. Dataset documentation is on the Training Data Transparency page.
License
Apache License 2.0. See LICENSE.
Contact
- Downloads last month
- 27