basikGPT-1-v1.0

basikGPT-1 v1.0

124M parameters (GPT-2 Small). v1.0 = 2.5B training tokens, not 2.5B parameters.

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("project-iconik/basikGPT-1-v1.0")
model = AutoModelForCausalLM.from_pretrained("project-iconik/basikGPT-1-v1.0")

Code and training recipe: github.com/project-iconik/basikGPT.
Whitepaper: EN · JA · KO.
Also available: basikGPT-1 v1.1 (same 124M, 5B tokens).

Model details

This Hub snapshot is a GPT2LMHeadModel export (safetensors + official GPT-2 tokenizer files). Optimizer state is not included.

Training data

• HuggingFaceFW/fineweb-edu (sample-10BT), first 2.5B tokens, sequential (no shuffle).
• FineWeb-Edu is released under ODC-By 1.0. Downstream use should respect that license and the original page licenses.

How it was trained

Single-GPU educational pretrain from random init. BF16, SDPA, batch 8 × 1024 × 8 grad accum (65,536 tokens/step), peak LR 6e-4 → min 6e-5, 2000-step warmup, cosine decay. Recipe and logs: the GitHub repo (docs/main_2p5b.md, runs/main_2p5b/).

Evaluation (same protocol as the repo)

Zero-shot English LM suite (english-lm-suite-v1) measured in-repo (not mixed with published paper numbers). Token counts and architectures are not matched across rows.

WG is acc_raw; other columns report the suite's primary metrics. Avg is the unweighted mean of these five primary metrics. n: HS 10,042 · LAMBADA 5,153 · PIQA 1,838 · WG 1,267 · ARC-E 2,376. Full protocol: benchmarks/REPORT.md. Methods and the full comparison: whitepaper.

v1.0 is the only checkpoint in this pair that beats official GPT-2 Small on ARC-Easy (+4.88pp). LAMBADA is still well below gpt2 (−11.35pp). WinoGrande is chance-level.

Intended use

Research and education: a small, reproducible GPT-2 Small trained on a documented public corpus. Not an instruction model. Not safe for open-ended production chat.

Limitations

• 124M / 2.5B tokens: weak on long-range last-word prediction and pronoun resolution.
• English-centric FineWeb-Edu; no instruction or preference tuning.
• May reproduce biases and factual errors from web text.

Citation

@software{basikgpt,
  title = {basikGPT},
  author = {basikGPT Contributors},
  url = {https://github.com/project-iconik/basikGPT},
  year = {2026}
}