Model Card for HuggingFaceFW/ablation-model-fineweb-edu

Model summary

This model is part of the 🍷 FineWeb ablations, detailed in this technical report. The model has 1.82B parameters, 2048 context length and uses Llama architecture with RoPE. It was trained on 350B tokens from FineWeb-Edu, tokenized using gpt2 tokenizer.

• Paper: 🍷 FineWeb: decanting the web for the finest text data at scale https://hf.co/spaces/HuggingFaceFW/blogpost-fineweb-v1
• License: Apache-2
• Languages: English

Use

Intended use

This model was trained on English web data and is not instruction-tuned, making it intended for text completion in English. It is important to note that the primary intended use case of this model is to compare its performance with other models trained under the same conditions. This model is not necessarily the best possible outcome achievable with the given dataset.

Generation

# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model = "HuggingFaceFW/ablation-model-fineweb-edu"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(model)
model = AutoModelForCausalLM.from_pretrained(model).to(device)

inputs = tokenizer.encode("Machine Learning is", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

Intermediate checkpoints (soon)

We are releasing intermediate checkpoints for this model at intervals of every 1000 training steps in separate branches. The naming convention is step-001000-2BT.

You can load a specific model revision with transformers using the argument revision:

model = AutoModelForCausalLM.from_pretrained("HuggingFaceFW/ablation-model-fineweb-edu", revision="step-001000-2BT")

You can access all the revisions for the models via the following code:

from huggingface_hub import list_repo_refs
out = list_repo_refs("HuggingFaceFW/ablation-model-fineweb-edu")
print([b.name for b in out.branches])

Training

Model

• Architecture: Llama model
• Pretraining steps: 167k
• Pretraining tokens: 350B
• Precision: bfloat16

Hardware

• GPUs: 64 H100
• Training time: 72 wall clock hours

Software

• nanotron for training
• datatrove for tokenization
• lighteval for evaluation

Evaluation

We used the same setup to evaluate all our ablation models with lighteval. To reproduce our numbers, make sure to follow the instruction here.

# download https://huggingface.co/datasets/HuggingFaceFW/fineweb/blob/main/lighteval_tasks.py and run:
accelerate launch --num_processes=1 lighteval/run_evals_accelerate.py --model_args="pretrained=HuggingFaceFW/ablation-model-fineweb-edu" \
    --custom_tasks "lighteval_tasks.py" --output_dir [OUTPUTPATH] --max_samples 1000 \ 
    --tasks "custom|hellaswag|0|1,custom|winogrande|0|1,custom|piqa|0|1,custom|siqa|0|1,custom|openbookqa|0|1,custom|arc:easy|0|1,custom|arc:challenge|0|1,custom|commonsense_qa|0|1,custom|mmlu:abstract_algebra|0|1,custom|mmlu:anatomy|0|1,custom|mmlu:astronomy|0|1,custom|mmlu:business_ethics|0|1,custom|mmlu:clinical_knowledge|0|1,custom|mmlu:college_biology|0|1,custom|mmlu:college_chemistry|0|1,custom|mmlu:college_computer_science|0|1,custom|mmlu:college_mathematics|0|1,custom|mmlu:college_medicine|0|1,custom|mmlu:college_physics|0|1,custom|mmlu:computer_security|0|1,custom|mmlu:conceptual_physics|0|1,custom|mmlu:econometrics|0|1,custom|mmlu:electrical_engineering|0|1,custom|mmlu:elementary_mathematics|0|1,custom|mmlu:formal_logic|0|1,custom|mmlu:global_facts|0|1,custom|mmlu:high_school_biology|0|1,custom|mmlu:high_school_chemistry|0|1,custom|mmlu:high_school_computer_science|0|1,custom|mmlu:high_school_european_history|0|1,custom|mmlu:high_school_geography|0|1,custom|mmlu:high_school_government_and_politics|0|1,custom|mmlu:high_school_macroeconomics|0|1,custom|mmlu:high_school_mathematics|0|1,custom|mmlu:high_school_microeconomics|0|1,custom|mmlu:high_school_physics|0|1,custom|mmlu:high_school_psychology|0|1,custom|mmlu:high_school_statistics|0|1,custom|mmlu:high_school_us_history|0|1,custom|mmlu:high_school_world_history|0|1,custom|mmlu:human_aging|0|1,custom|mmlu:human_sexuality|0|1,custom|mmlu:international_law|0|1,custom|mmlu:jurisprudence|0|1,custom|mmlu:logical_fallacies|0|1,custom|mmlu:machine_learning|0|1,custom|mmlu:management|0|1,custom|mmlu:marketing|0|1,custom|mmlu:medical_genetics|0|1,custom|mmlu:miscellaneous|0|1,custom|mmlu:moral_disputes|0|1,custom|mmlu:moral_scenarios|0|1,custom|mmlu:nutrition|0|1,custom|mmlu:philosophy|0|1,custom|mmlu:prehistory|0|1,custom|mmlu:professional_accounting|0|1,custom|mmlu:professional_law|0|1,custom|mmlu:professional_medicine|0|1,custom|mmlu:professional_psychology|0|1,custom|mmlu:public_relations|0|1,custom|mmlu:security_studies|0|1,custom|mmlu:sociology|0|1,custom|mmlu:us_foreign_policy|0|1,custom|mmlu:virology|0|1,custom|mmlu:world_religions|0|1"

In particular the MMLU prompts are slightly different from those in lm-evaluation-harness and the Open LLM Leaderboard, more in this blogpost. We use prompt templates that provide better signal for small and non instruction tuned models.

Limitations

This model was predominantly trained on English data, potentially limiting its performance in other languages. Furthermore, the model's behavior is influenced by the quality and diversity of its training data, which may include biases and harmful content.