wiki-model-GGUF / README.md
Lucasgitguy-co's picture
Add GGUF conversion (tested with llama.cpp build 12359)
1fd4211 verified
|
Raw History Blame Contribute Delete
2.43 kB
metadata
license: apache-2.0
base_model: Lucasgitguy-co/wiki-model
base_model_relation: quantized
library_name: gguf
pipeline_tag: text-generation
language: en
tags:
  - gguf
  - llama.cpp
  - wikipedia
  - facts

wiki-model GGUF

GGUF files of Lucasgitguy-co/wiki-model, a SmolLM2-135M fine-tune that writes short facts, answers questions and fixes false statements. They run in llama.cpp and in apps built on it.

File Type Size (GB)
wiki-model-F16.gguf F16, same output as the original 0.271
wiki-model-Q8_0.gguf Q8_0 0.145
wiki-model-Q4_K_M.gguf Q4_K_M 0.105

Use

The model has no chat template. It reads tags, the same as the Python package:

Task Prompt
Fact about a topic <fact> Amsterdam </fact>
Question <q> Who wrote Hamlet? </q>
Fix a false claim <fix> The Eiffel Tower is in Berlin. </fix>
With your own source text <ctx> some text </ctx> <q> question </q>
hf download Lucasgitguy-co/wiki-model-GGUF wiki-model-Q8_0.gguf --local-dir .
llama-completion -m wiki-model-Q8_0.gguf -no-cnv --temp 0 -n 100 -p "<q> Who wrote Hamlet? </q>"

The --grounded mode in the Python package (a BM25 index of Wikipedia) is not part of these files. Without context the model's facts are often wrong: in the source evaluation, 5-12% of model-only facts were supported. Paste source text in <ctx> to get grounded answers. The Python CLI also blocks repeated 4-grams inside the answer, and llama.cpp doesn't do that by default.

How these were made and tested

  • Converted with llama.cpp convert_hf_to_gguf.py (checkout 1b6935b). The source was saved with transformers 5, so two fields were rewritten for the converter: rope_parameters.rope_theta (100000) was copied to rope_theta, and the extra_special_tokens list became additional_special_tokens. The weights were not changed.
  • Tested 2026-11-16 with llama.cpp 1.0.0 (build 12359):
    • Token ids match the transformers tokenizer.
    • Greedy output of the F16 file is identical to transformers (float32) on 4 prompts, 40 tokens each.
    • Q8_0 and Q4_K_M were only spot-checked: they gave the same answer on one prompt.

Licence

Apache-2.0, the same as the source model. Wikipedia-derived training data is CC BY-SA 4.0. See the source card.