wiki-model-tiny GGUF
GGUF files of Lucasgitguy-co/wiki-model-tiny, a 12.5M-parameter model trained from scratch. It runs in llama.cpp as a normal llama model.
| File | Type | Size (GB) |
|---|---|---|
wiki-model-tiny-F16.gguf |
F16, same output as the original | 0.025 |
wiki-model-tiny-Q8_0.gguf |
Q8_0 | 0.014 |
There are no K-quants: the model width (320) isn't a multiple of 256.
Use
The prompt tags are <fact> topic </fact>, <q> question </q>, <fix> claim </fix>, and an optional <ctx> text </ctx> in front. The context window is 384 tokens.
hf download Lucasgitguy-co/wiki-model-tiny-GGUF wiki-model-tiny-F16.gguf --local-dir .
llama-completion -m wiki-model-tiny-F16.gguf -no-cnv --temp 0 -n 80 -c 384 \
-p "<ctx> Mount Everest is Earth's highest mountain, in the Himalayas. </ctx> <q> Where is Mount Everest? </q>"
The Python package's --grounded BM25 index isn't included, so supply the context yourself. Without context this model mostly makes things up.
How these were made and tested
- The original is plain PyTorch, so
convert_tiny.pywrites the GGUF directly with gguf-py:- The fused
qkvandgate_upweights are split. - RoPE is kept as is, because the model already uses llama.cpp's interleaved layout.
- The byte-level BPE tokenizer is stored as
gpt2with thegpt-2pre-tokenizer.
- The fused
- Tested 2026-11-16 with llama.cpp 1.0.0 (build 12359):
- Token ids match the original tokenizer on all three prompt formats.
- Greedy output (F32 and F16) is identical to the PyTorch model on 4 prompts, up to the end-of-answer token.
- Q8_0 was spot-checked on one prompt.
Licence
Apache-2.0, the same as the source model. Wikipedia-derived training data is CC BY-SA 4.0.