collinear-ai/math_reasoning_phi_c1

See axolotl config

axolotl version: 0.5.0

Collinear Curator 1:

This is an open-source fine-tuned reasoning adapter of microsoft/Phi-3.5-mini-instruct, transformed into a math reasoning model using data curated from collinear-ai/R1-Distill-SFT-Curated. It achieves the following results on the evaluation set:

• Loss: 0.3203

Model description

This model is a LoRA adaptor and for best results merge it with base model microsoft/Phi-3.5-mini-instruct before use.

Training and evaluation data

• Training data: collinear-ai/R1-Distill-SFT-Curated
• Evaluation data: HuggingFaceH4/MATH-500

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 50
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0003	1	0.6714
0.337	0.3335	1243	0.3361
0.3248	0.6669	2486	0.3203

Evaluation Results on Math500

The following figure shows the accuracy and the speedup of Collinear Curators C1 and C2 when compared to training on unfiltered dataset.

Framework versions

• PEFT 0.13.2
• Transformers 4.46.1
• Pytorch 2.3.1+cu121
• Datasets 3.0.1
• Tokenizers 0.20.3