Downloads · 30 days
71
45% of all-time downloads
leeaandrob/mythos-d13x1872-r373-ptbr-gguf
mythos-d13x1872-r373-ptbr-gguf is a machine learning model from leeaandrob. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
A small Portuguese language model trained from scratch, quantized to ternary Q20 and served by the NeuroGrid engine. This repo holds the deployed checkpoint: the d13 configuration, nembd 1872, after supervised fine-tu…
Downloads · 30 days
71
45% of all-time downloads
All-time downloads
157
Public
Repo size
2.5 GB
Likes
0
Public
Click a slice to open those files.
.gguf2.5 GB · 100%
From the Hugging Face model README
A small Portuguese language model trained from scratch, quantized to ternary
Q2_0 and served by the NeuroGrid engine.
This repo holds the deployed checkpoint: the d13 configuration, n_embd 1872,
after supervised fine-tuning.
| File (recommended) | mythos_d13_final_q2_0_f16.gguf (898 MB) |
| File (original) | mythos_d13_final_q2_0.gguf (1.65 GB) |
| Tokenizer | mythos_tokenizer_d13.json, in this repo, required |
| Language | Portuguese (pt-BR) |
| Quantization | ternary Q2_0 body; embeddings, value embeddings, adapter and lm_head in F16 (recommended file) or F32 (original file) |
| val_bpb | 1.173 at pre-training step 13800, 0.709 after SFT step 213 |
Both files hold the same model. The ternary layers are identical; only the floating-point tables differ. The source checkpoint is bf16, so storing those tables in F16 instead of F32 drops no information, and the two files produce the same logits.
The F16 file is half the size, loads without copying the tables on the CPU
backend, and uses less VRAM on CUDA. It needs a mythos-serve built from the
NeuroGrid engine after 2026-07-29; older builds read only the F32 file.
mythos-serve \
-gguf mythos_d13_final_q2_0_f16.gguf \
-tokenizer mythos_tokenizer_d13.json \
-backend cuda \
-port 8095
-backend cpu and -backend metal also work; the tokenizer file is not
optional, since the GGUF does not carry it.
It is a small model, and it is used as one: PT-BR generation behind a
retrieval layer, where the retrieved passages carry the facts and the model
supplies the language. In the STJ jurisprudence RAG stack it is the
alternative generator behind mythos-stj, sitting next to the 27B behind
bonsai-stj, so the two can be compared on the same retrieved context.
Do not use it as a standalone knowledge source. It has no useful factual recall at this size and will produce fluent Portuguese that is wrong.