Text Generation
Transformers
Safetensors
Polish
gpt2
polish
nanogpt
tokenizer-ablation
research
Eval Results (legacy)
text-generation-inference
Instructions to use KateMajzel/GoLLeM-45M-PL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use KateMajzel/GoLLeM-45M-PL with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="KateMajzel/GoLLeM-45M-PL")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("KateMajzel/GoLLeM-45M-PL") model = AutoModelForCausalLM.from_pretrained("KateMajzel/GoLLeM-45M-PL", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use KateMajzel/GoLLeM-45M-PL with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "KateMajzel/GoLLeM-45M-PL" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KateMajzel/GoLLeM-45M-PL", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/KateMajzel/GoLLeM-45M-PL
- SGLang
How to use KateMajzel/GoLLeM-45M-PL with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "KateMajzel/GoLLeM-45M-PL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KateMajzel/GoLLeM-45M-PL", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "KateMajzel/GoLLeM-45M-PL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KateMajzel/GoLLeM-45M-PL", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use KateMajzel/GoLLeM-45M-PL with Docker Model Runner:
docker model run hf.co/KateMajzel/GoLLeM-45M-PL
|
Download README.md from KateMajzel/GoLLeM-45M-PL: direct link, hf CLI and curl.
- Browser
- Download file 11.9 kB
-
https://huggingface.co/KateMajzel/GoLLeM-45M-PL/resolve/main/README.md
- Command line
-
hf download hf://KateMajzel/GoLLeM-45M-PL/README.md
-
curl -L -o README.md https://huggingface.co/KateMajzel/GoLLeM-45M-PL/resolve/main/README.md
11.9 kB
| language: | |
| - pl | |
| license: mit | |
| library_name: transformers | |
| pipeline_tag: text-generation | |
| tags: | |
| - polish | |
| - gpt2 | |
| - nanogpt | |
| - tokenizer-ablation | |
| - research | |
| model-index: | |
| - name: GoLLeM-45M-PL | |
| results: | |
| - task: | |
| type: text-classification | |
| name: Sentiment Classification (zero-shot, log-likelihood) | |
| dataset: | |
| type: allegro/klej-polemo2-in | |
| name: PolEmo2.0-IN | |
| split: test | |
| config: polemo2-in | |
| metrics: | |
| - type: accuracy | |
| value: 47.2 | |
| name: Accuracy (PMI-normalized) | |
| verified: false | |
| - type: accuracy | |
| value: 40.0 | |
| name: Majority-class baseline | |
| verified: false | |
| - task: | |
| type: text-classification | |
| name: Topic Classification (zero-shot, log-likelihood) | |
| dataset: | |
| type: sdadas/8tags | |
| name: 8Tags | |
| split: test | |
| metrics: | |
| - type: accuracy | |
| value: 31.5 | |
| name: Accuracy (PMI-normalized) | |
| verified: false | |
| - type: accuracy | |
| value: 16.5 | |
| name: Majority-class baseline | |
| verified: false | |
| - task: | |
| type: text-generation | |
| name: Language Modeling | |
| dataset: | |
| type: speakleash | |
| name: SpeakLeash held-out (private, 1999 docs) | |
| metrics: | |
| - type: bits_per_byte | |
| value: 1.2114 | |
| name: BPB (mean of 3 seeds) | |
| verified: false | |
| - type: bits_per_byte | |
| value: 0.0100 | |
| name: BPB std. dev. across seeds | |
| verified: false | |
| # GoLLeM-45M-PL | |
| A GPT-2 model (42.5M parameters) trained from scratch on 2.96 GB of Polish text, | |
| with a dedicated Polish BPE tokenizer (32,768 entries). | |
| The model is an **artifact of a research experiment**, not a practical tool. | |
| The question: *does a dedicated Polish tokenizer produce a better model than the GPT-2 | |
| tokenizer under the same training budget?* The answer: **it is not better — it is | |
| comparable at 2.5× lower cost.** | |
| Code, data and full methodology: https://github.com/KateMajzel/gollem-pl | |
| Mirror: [SlayerLab/GoLLeM-45M-PL](https://huggingface.co/SlayerLab/GoLLeM-45M-PL). | |
| ## Usage | |
| ```python | |
| from transformers import pipeline | |
| pipe = pipeline("text-generation", model="KateMajzel/GoLLeM-45M-PL") | |
| print(pipe("W lesie mieszkał mały", max_new_tokens=50)[0]["generated_text"]) | |
| ``` | |
| Default generation parameters: `do_sample=True`, `temperature=0.8`, `top_p=0.9`, | |
| `repetition_penalty=1.15`. Greedy decoding falls into repetition loops — typical for | |
| models of this size. The 199 unused filler tokens (32,569–32,767) are blocked via | |
| `bad_words_ids`. | |
| ## Results | |
| BPB on the private held-out set (1,999 documents, an identical 2,767,440 bytes for | |
| each model; means over 3 seeds for R1 and R2a): | |
| | run | tokenizer | budget | BPB ↓ | sd | parameters | time | | |
| |---|---|---|---|---|---|---| | |
| | **R1** (this model) | Polish, 32,768 | 2.96 GB | 1.2114 | 0.0100 | 42.5M | **49.7 min** | | |
| | R2a | GPT-2, 50,257 | 2.96 GB — same bytes | 1.1946 | 0.0035 | 51.5M | 122.4 min | | |
| | R2b | GPT-2, 50,257 | 2,970 steps — same tokens | 1.2756 | (n=1) | 51.5M | 62.4 min | | |
| | R3 | GPT-2 zero-shot | — | 2.9555 | (n=1) | 124M | — | | |
| Zero-shot benchmarks (log-likelihood with domain PMI normalization): | |
| | task | R1 | R2a | GPT-2 | majority class | random | | |
| |---|---|---|---|---|---| | |
| | PolEmo2-IN | 47.2% | 43.8% | 20.8% | 40.0% | 25.0% | | |
| | 8Tags | 31.5% | 29.8% | 17.8% | 16.5% | 12.5% | | |
| ### Interpretation | |
| Four independent measurements (BPB, PolEmo2, 8Tags, the private held-out set) show no | |
| significant advantage for either tokenizer under an equal text budget, and their | |
| directions are inconsistent. The decisive observation: **the BPB difference between two | |
| seeds of the same model (0.0186) exceeds the difference between models with different | |
| tokenizers (0.0124).** | |
| The advantage lies in cost. Under a matched compute budget (R1 vs R2b) the Polish | |
| tokenizer wins by 5.30% BPB — 6.4 standard deviations. The Bielik v3 PL team reached | |
| the same conclusion independently at 11B scale (arXiv 2604.10799). | |
| ### A note on perplexity | |
| PPL/token is 30.7 for R1 and 5.8 for R2a, which would suggest a fivefold advantage for | |
| GPT-2. This is an artifact: models with different tokenizers predict units of different | |
| difficulty. **Perplexity is not comparable between them.** The only correct metric is | |
| bits-per-byte. | |
| ## Tokenizer | |
| Byte-level BPE, 32,568 entries + 200 special. 27.9% of entries contain Polish | |
| diacritics, lossless round-trip, zero tokens unreachable through merges. | |
| | | this tokenizer | GPT-2 | | |
| |---|---|---| | |
| | bytes/token (corpus) | 4.050 | 2.066 | | |
| | bytes/token (held-out) | 4.103 | 2.134 | | |
| | density | **1.96×** | 1.00× | | |
| ## Training data | |
| 2.96 GB, 803,177 documents, [SpeakLeash](https://speakleash.org/): 37% encyclopedia and | |
| literature, 22% forums, 27% web, 11% journalism and science, 3% official/administrative | |
| texts. | |
| Deliberately excluded: translations and subtitles (translationese), synthetic text, song | |
| lyrics (copyright), raw Common Crawl. Dataset licensing is handled on the SpeakLeash side. | |
| ## Architecture | |
| 8 layers × 8 heads × 512 dimensions, 1,024-token context, 42.5M parameters | |
| (25.7M non-embedding), 731M training tokens (17.2 per parameter), | |
| AdamW lr 1e-3 → 1e-4, bfloat16, 1 epoch, 1× RTX 5080. | |
| ## Limitations | |
| - **Scale.** The model produces correct Polish at the sentence level, but hallucinates | |
| facts and loses coherence after a few sentences. An example from generation: | |
| *"wieś znalazła się w powiecie częstochowskim, gminie Łęczna"* ("the village was located | |
| in Częstochowa County, Łęczna municipality") — both names exist, the combination is false. | |
| - **Base model.** Text continuation only; no instruction tuning and no safety filtering. | |
| - **Internet forum data.** The model may reproduce the biases and language present there. | |
| It is not suitable for end-user applications — children in particular — without human | |
| oversight. | |
| - **A single scale and a single configuration.** The results apply to 42M parameters and | |
| 3 GB of data. | |
| - **Register depends on the data mix.** Given a children's fairy-tale prompt, the model | |
| responds in the language of forums or encyclopedia entries. | |
| - The corpus contains Windows line endings (`\r\n`), which the model reproduces. | |
| ## Reproducibility | |
| Code, configurations, evaluation scripts and a description of five methodological | |
| pitfalls: https://github.com/KateMajzel/gollem-pl | |
| Based on [nanoGPT](https://github.com/karpathy/nanoGPT) (MIT, Andrej Karpathy). | |
| ## Citation | |
| ```bibtex | |
| @misc{gollem45mpl, | |
| title = {GoLLeM-45M-PL: a tokenizer ablation for a Polish language model}, | |
| author = {Majzel-Pośpiech, Katarzyna}, | |
| year = {2026}, | |
| url = {https://huggingface.co/KateMajzel/GoLLeM-45M-PL} | |
| } | |
| ``` | |
| --- | |
| # GoLLeM-45M-PL | |
| Model GPT-2 (42,5 mln parametrów) wytrenowany od zera na 2,96 GB polskiego tekstu, | |
| z dedykowanym polskim tokenizerem BPE (32 768 pozycji). | |
| Model jest **artefaktem eksperymentu badawczego**, nie narzędziem użytkowym. | |
| Pytanie: *czy dedykowany tokenizer polski daje lepszy model niż tokenizer GPT-2 przy | |
| tym samym budżecie treningowym?* Odpowiedź: **nie daje lepszego — daje porównywalny | |
| 2,5× taniej.** | |
| Kod, dane i pełna metodologia: https://github.com/KateMajzel/gollem-pl | |
| Kopia lustrzana: [SlayerLab/GoLLeM-45M-PL](https://huggingface.co/SlayerLab/GoLLeM-45M-PL). | |
| ## Użycie | |
| ```python | |
| from transformers import pipeline | |
| pipe = pipeline("text-generation", model="KateMajzel/GoLLeM-45M-PL") | |
| print(pipe("W lesie mieszkał mały", max_new_tokens=50)[0]["generated_text"]) | |
| ``` | |
| Domyślne parametry generacji: `do_sample=True`, `temperature=0.8`, `top_p=0.9`, | |
| `repetition_penalty=1.15`. Dekodowanie zachłanne wpada w pętle powtórzeń — typowe dla | |
| modeli tej wielkości. 199 nieużywanych tokenów wypełniających (32 569–32 767) | |
| zablokowano przez `bad_words_ids`. | |
| ## Wyniki | |
| BPB na prywatnym held-oucie (1 999 dokumentów, identyczne 2 767 440 bajtów dla | |
| każdego modelu; średnie z 3 ziaren dla R1 i R2a): | |
| | przebieg | tokenizer | budżet | BPB ↓ | sd | parametry | czas | | |
| |---|---|---|---|---|---|---| | |
| | **R1** (ten model) | polski, 32 768 | 2,96 GB | 1,2114 | 0,0100 | 42,5 M | **49,7 min** | | |
| | R2a | GPT-2, 50 257 | 2,96 GB — te same bajty | 1,1946 | 0,0035 | 51,5 M | 122,4 min | | |
| | R2b | GPT-2, 50 257 | 2 970 kroków — te same tokeny | 1,2756 | (n=1) | 51,5 M | 62,4 min | | |
| | R3 | GPT-2 zero-shot | — | 2,9555 | (n=1) | 124 M | — | | |
| Benchmarki zero-shot (log-likelihood z normalizacją PMI domenową): | |
| | zadanie | R1 | R2a | GPT-2 | klasa większościowa | losowo | | |
| |---|---|---|---|---|---| | |
| | PolEmo2-IN | 47,2% | 43,8% | 20,8% | 40,0% | 25,0% | | |
| | 8Tags | 31,5% | 29,8% | 17,8% | 16,5% | 12,5% | | |
| ### Interpretacja | |
| Cztery niezależne pomiary (BPB, PolEmo2, 8Tags, prywatny held-out) nie wykazują | |
| istotnej przewagi żadnego tokenizera przy równym budżecie tekstu, a ich kierunki są | |
| niespójne. Rozstrzygająca obserwacja: **różnica BPB między dwoma ziarnami tego samego | |
| modelu (0,0186) przewyższa różnicę między modelami o odmiennych tokenizerach (0,0124).** | |
| Przewaga leży w koszcie. Przy wyrównanym budżecie obliczeniowym (R1 vs R2b) tokenizer | |
| polski wygrywa o 5,30% BPB — 6,4 odchylenia. Ten sam wniosek uzyskał niezależnie zespół | |
| Bielika v3 PL na skali 11B (arXiv 2604.10799). | |
| ### Uwaga o perplexity | |
| PPL/token wynosi 30,7 dla R1 i 5,8 dla R2a, co sugerowałoby pięciokrotną przewagę | |
| GPT-2. To artefakt: modele o różnych tokenizerach przewidują jednostki różnej trudności. | |
| **Perplexity jest między nimi nieporównywalna.** Jedyną poprawną metryką jest | |
| bits-per-byte. | |
| ## Tokenizer | |
| Byte-level BPE, 32 568 pozycji + 200 specjalnych. 27,9% pozycji z polskimi | |
| diakrytykami, round-trip bezstratny, zero tokenów nieosiągalnych przez merge. | |
| | | ten tokenizer | GPT-2 | | |
| |---|---|---| | |
| | bajty/token (korpus) | 4,050 | 2,066 | | |
| | bajty/token (held-out) | 4,103 | 2,134 | | |
| | gęstość | **1,96×** | 1,00× | | |
| ## Dane treningowe | |
| 2,96 GB, 803 177 dokumentów, [SpeakLeash](https://speakleash.org/): 37% encyklopedia | |
| i literatura, 22% fora, 27% web, 11% publicystyka i nauka, 3% teksty urzędowe. | |
| Celowo wykluczone: tłumaczenia i napisy (translationese), tekst syntetyczny, teksty | |
| piosenek (prawa autorskie), surowy Common Crawl. Licencje zbiorów po stronie SpeakLeash. | |
| ## Architektura | |
| 8 warstw × 8 głowic × 512 wymiarów, kontekst 1 024 tokeny, 42,5 M parametrów | |
| (25,7 M niezanurzeniowych), 731 mln tokenów treningowych (17,2 na parametr), | |
| AdamW lr 1e-3 → 1e-4, bfloat16, 1 epoka, 1× RTX 5080. | |
| ## Ograniczenia | |
| - **Skala.** Model produkuje poprawną polszczyznę na poziomie zdania, ale halucynuje | |
| fakty i traci spójność po kilku zdaniach. Przykład z generacji: *„wieś znalazła się | |
| w powiecie częstochowskim, gminie Łęczna"* — obie nazwy istnieją, zestawienie jest | |
| fałszywe. | |
| - **Model bazowy.** Wyłącznie kontynuacja tekstu; bez dostrajania instrukcyjnego | |
| i bez filtrowania bezpieczeństwa. | |
| - **Dane z forów internetowych.** Model może odtwarzać obecne tam uprzedzenia i język. | |
| Nie nadaje się do zastosowań z udziałem użytkownika końcowego — w szczególności | |
| dzieci — bez nadzoru człowieka. | |
| - **Jedna skala i jedna konfiguracja.** Wyniki dotyczą 42 M parametrów i 3 GB danych. | |
| - **Rejestr zależny od miksu danych.** Przy prompcie z bajki dla dzieci model | |
| odpowiada językiem forów lub haseł encyklopedycznych. | |
| - Korpus zawiera windowsowe końce linii (`\r\n`), które model odtwarza. | |
| ## Odtwarzalność | |
| Kod, konfiguracje, skrypty ewaluacyjne i opis pięciu pułapek metodologicznych: | |
| https://github.com/KateMajzel/gollem-pl | |
| Bazuje na [nanoGPT](https://github.com/karpathy/nanoGPT) (MIT, Andrej Karpathy). | |
| ## Cytowanie | |
| ```bibtex | |
| @misc{gollem45mpl, | |
| title = {GoLLeM-45M-PL: ablacja tokenizera dla polskiego modelu językowego}, | |
| author = {Majzel-Pośpiech, Katarzyna}, | |
| year = {2026}, | |
| url = {https://huggingface.co/KateMajzel/GoLLeM-45M-PL} | |
| } | |
| ``` | |