fineweb-mixtral-edu-score

This model is a fine-tuned version of intfloat/multilingual-e5-base on an transferred from English dataset. It achieves the following results on the evaluation set:

• Loss: 0.0636
• Precision: 0.9335
• Recall: 0.9445
• F1 Macro: 0.9386
• Accuracy: 0.9446

Model description

This model measure educational value of the given text for humans, as labelled by Mixtral model.

Intended uses & limitations

Data filtering and evaluation of pretraining data at scale.

Training and evaluation data

Take a look at https://github.com/lapa-llm/lapa-llm/blob/main/pretraining/quality-classifiers/fineweb_hf.py

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 128
• seed: 0
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 256
• total_eval_batch_size: 1024
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1 Macro	Accuracy
No log	0	0	0.2341	0.4596	0.4913	0.4228	0.6411
0.0758	0.7968	200	0.0720	0.9193	0.9190	0.9191	0.9280
0.0682	1.5936	400	0.0664	0.9296	0.9352	0.9323	0.9393
0.0657	2.3904	600	0.0644	0.9323	0.9437	0.9375	0.9436
0.0648	3.1873	800	0.0642	0.6204	0.6303	0.6248	0.9431
0.064	3.9841	1000	0.0636	0.9335	0.9445	0.9386	0.9446

Framework versions

• Transformers 4.56.1
• Pytorch 2.6.0a0+ecf3bae40a.nv25.01
• Datasets 4.0.0
• Tokenizers 0.22.0