Another experiment in the line of loyal-piano-m7.
Steps taken to produce this model:
- Train loyal-piano-m7
- cDPO with HuggingFaceH4/ultrafeedback_binarized to produce loyal-piano-m7-cdpo
- Train another model with different sampling of the same source datasets as loyal-piano, let's call it servile-harpsichord
- cDPO servile-harpsichord with allenai/ultrafeedback_binarized_cleaned, Intel/orca_dpo_pairs, and a helpfulness-only version of PKU-Alignment/PKU-SafeRLHF
- TIES merge several checkpoints of servile-harpsichord-cdpo with loyal-piano-m7-cdpo
Local benchmarks show the result to be better than any of the individual components. Let's see if that holds up!
Trained using the Alpaca prompt format.
Configuration for final merge:
models:
- model: chargoddard/loyal-piano-m7-cdpo
parameters:
density: 1.0
weight: 1.0
- model: /home/ubuntu/servile-harpsichord-cdpo/checkpoint-4186
parameters:
weight: 0.1
- model: /home/ubuntu/servile-harpsichord-cdpo/checkpoint-5796
parameters:
weight: 0.2
- model: /home/ubuntu/servile-harpsichord-cdpo/checkpoint-6118
parameters:
weight: 0.3
- model: /home/ubuntu/servile-harpsichord-cdpo/final
parameters:
weight: 0.4
merge_method: ties
base_model: mistralai/Mistral-7B-v0.1
dtype: bfloat16
parameters:
density: 0.4
normalize: true
int8_mask: true