---

license: apache-2.0

library_name: transformers

pipeline_tag: text-generation

base_model: TurkuNLP/gpt3-finnish-small

tags:

• gguf
• llama.cpp
• text-generation
• finnish
• suomi
• bloom
• 186m
• small-model
• quantization
• q4_k_m
• q8_0

---

GPT3 Finnish Small GGUF

<div align="center">

<img alt="Model" src="https://img.shields.io/badge/model-gpt3--finnish--small-8A2BE2?style=for-the-badge">

<img alt="GGUF formats" src="https://img.shields.io/badge/GGUF-F16%20%7C%20Q8_0%20%7C%20Q4_K_M-FFD21E?style=for-the-badge">

<img alt="Architecture" src="https://img.shields.io/badge/arch-BLOOM-00A6A6?style=for-the-badge">

<img alt="Parameters" src="https://img.shields.io/badge/params-186M-16A34A?style=for-the-badge">

<img alt="Language" src="https://img.shields.io/badge/language-Finnish-0069B4?style=for-the-badge">

<img alt="License" src="https://img.shields.io/badge/license-Apache--2.0-7C3AED?style=for-the-badge">

</div>

Source model · HF release · Build hub

Native llama.cpp GGUF release for the 186M-parameter Finnish BLOOM base model. The source config omits seq_length; the build restores the tokenizer's 2,048-token limit before conversion.

Formats

| File | Status | Purpose |

|---|---|---|

| gpt3-finnish-small-F16.gguf | Published | Reference quality |

| gpt3-finnish-small-Q8_0.gguf | Published | Higher-quality compact format |

| gpt3-finnish-small-Q4_K_M.gguf | Published | Smallest release format |

Verified outputs — read before using

Verbatim gpt3-finnish-small-Q4_K_M.gguf completions, --temp 0, 24 new tokens, prompt form Kysymys: ...\nVastaus:. All eight rows are published, including the wrong ones.

Correct answers

| Kysymys | Model answer |

|---|---|

| Mikä on Suomen pääkaupunki? | Helsinki on Suomen pääkaupunki. (then repeats the same sentence to fill the token budget) |

| Mikä on Ranskan pääkaupunki? | Ranskan pääkaupunki on Pariisi. |

| Onko Helsinki Pohjoismaiden pääkaupunki? | Helsinki on Pohjoismaiden pääkaupunki. |

Wrong or degenerate answers

| Kysymys | Model answer |

|---|---|

| Kuinka monta päivää viikossa on? | Jos olet raskaana, sinun on oltava vähintään kaksi viikkoa viikossa. |

| Mikä planeetta on lähinnä aurinkoa? | Aurinko on aurinko, mutta se on myös planeetta. |

| Mikä on veden kemiallinen merkki? | Veden kemiallinen merkki on vesi. |

| Mikä on maailman suurin valtameri? | Maailman suurin valtameri on maailman suurin valtameri. |

| Mikä on Suomen kieli? | Suomen kieli on yksi maailman vanhimmista kielistä. (no answer to the question) |

3 of 8 factual questions correct, and every completion repeats itself. This is a 186M base model, so it is useful for Finnish perplexity baselines, tokenizer and BLOOM-architecture GGUF testing, and continued-pretraining experiments. It is not usable as a question-answering model, and quantisation is not the limiting factor here: F16 fails the same way.

Validation

Finnish Wikipedia evaluation, 8 chunks of 512 tokens. Lower perplexity is better.

| Format | PPL | Ratio to F16 |

|---|---:|---:|

| F16 | 44.4931 | Baseline |

| Q8_0 | 44.5125 | 1.0004 |

| Q4_K_M | 45.2245 | 1.0164 |

This is a conversion release, not an instruction or chat model.

Build

python -m pip install -r requirements-build.txt
python build_gguf.py --model-id TurkuNLP/gpt3-finnish-small

The builder pins llama.cpp commit 6b790a9c291b5d7af3312bbf9f0c558aa023b13e and the upstream model revision. It does not upload or overwrite this repository.

License

Apache-2.0. See the upstream model card and LICENSE for source attribution.