---
datasets: HuggingFaceFW/fineweb-edu
widget:
- example_title: Example interaction
  text: During photosynthesis in green plants
inference:
  parameters:
    repetition_penalty: 1.3
language:
- en
library_name: transformers
license: mit
---

# Model Card for gpt2-124M-edu-fineweb-10B

A 124M parameter GPT2 model trained with the 10B fineweb-edu dataset using [https://github.com/karpathy/llm.c](https://github.com/karpathy/llm.c)

Training took 20 hours on a single 4090 GPU (limited to 350W), giving the following graphs:

![gpt2-124M-edu-fineweb-10B](https://huggingface.co/rhysjones/gpt2-124M-edu-fineweb-10B/resolve/main/graph.png)

## Training

The training parameters where:
```
./train_gpt2cu \
    -i "dev/data/edu_fineweb10B/edu_fineweb_train_*.bin" \
    -j "dev/data/edu_fineweb10B/edu_fineweb_val_*.bin" \
    -o log124M \
    -e "d12" \
    -b 56 -t 1024 \
    -d 458752 \
    -r 1 \
    -z 1 \
    -c 0.1 \
    -l 0.002 \
    -q 0.0 \
    -u 700 \
    -n 5000 \
    -v 250 -s 20000 \
    -h 1
```

The model has had no further finetuning.

## Evaluation

Evals using [Eleuther AI Harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463)
as described in the [open_llm_leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
and comparing with those published for [openai-community/gpt2](https://huggingface.co/openai-community/gpt2)

![gpt2-124M-edu-fineweb-10B Evals](https://huggingface.co/rhysjones/gpt2-124M-edu-fineweb-10B/resolve/main/evals_124M.png)

| Eval Test | Score |
| --------- | ----- |
| arc_challenge (25 shot) | 24.49 |
| gsm8k (5 shot) | 0.08 |
| hellaswag (10 shot) | 32.64 |
| mmlu (5 shot) | 26.06 |
| truthfulqa (0 shot) | 42.45 |
| winogrande (5 shot) | 52.17 |
| **Overall Score** | **29.65** |