mistral7b-sharded-finetune-bn22

This model is a fine-tuned version of filipealmeida/Mistral-7B-Instruct-v0.1-sharded on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.1132

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• training_steps: 2500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
1.8338	0.16	100	1.4732
1.3163	0.33	200	1.2741
1.2292	0.49	300	1.2399
1.1665	0.65	400	1.2146
1.1597	0.82	500	1.1913
1.1025	0.98	600	1.1699
1.03	1.14	700	1.1546
1.0461	1.31	800	1.1491
1.0149	1.47	900	1.1334
0.9989	1.63	1000	1.1270
1.0385	1.79	1100	1.1184
1.0051	1.96	1200	1.1102
0.9365	2.12	1300	1.1210
0.8931	2.28	1400	1.1105
0.9094	2.45	1500	1.1095
0.8989	2.61	1600	1.1079
0.9027	2.77	1700	1.1043
0.9007	2.94	1800	1.1010
0.8666	3.1	1900	1.1111
0.8259	3.26	2000	1.1128
0.8288	3.43	2100	1.1153
0.8223	3.59	2200	1.1133
0.7891	3.75	2300	1.1132

Framework versions

• PEFT 0.7.0
• Transformers 4.36.0.dev0
• Pytorch 2.0.0
• Datasets 2.1.0
• Tokenizers 0.15.0