Zangs3011's picture
Update README.md
e5d36e3
metadata
library_name: peft
tags:
  - meta-llama
  - code
  - instruct
  - WizardLM
  - Mistral-7B-v0.1
datasets:
  - WizardLM/WizardLM_evol_instruct_70k
base_model: mistralai/Mistral-7B-v0.1
license: apache-2.0

Finetuning Overview:

Model Used: mistralai/Mistral-7B-v0.1
Dataset: WizardLM/WizardLM_evol_instruct_70k

Dataset Insights:

The WizardLM/WizardLM_evol_instruct_70k dataset, tailored specifically for enhancing interactive capabilities, it was developed using EVOL-Instruct method.Which will basically enhance a smaller dataset, with tougher quesitons for the LLM to perform

Finetuning Details:

With the utilization of MonsterAPI's LLM finetuner, this finetuning:

  • Was achieved with great cost-effectiveness.
  • Completed in a total duration of 5hrs 18mins for 1 epoch using an A6000 48GB GPU.
  • Costed $10 for the entire epoch.

Hyperparameters & Additional Details:

  • Epochs: 1
  • Cost Per Epoch: $10
  • Total Finetuning Cost: $10
  • Model Path: mistralai/Mistral-7B-v0.1
  • Learning Rate: 0.0002
  • Data Split: 90% train 10% validation
  • Gradient Accumulation Steps: 4

Prompt Structure

### INSTRUCTION:
[instruction]

### RESPONSE:
[output]

Training loss : training loss


Benchmark Results

ARC HELLSWAG TRUTHFULMQ Benchmark comparison

ARC (arc_challenge, acc_norm)   0.5543
HellaSwag (hellaswag, acc_norm) 0.7979
TruthfulQA (truthfulqa_mc2)     0.4781

license: apache-2.0