codellama-7b-humaneval-java-fim

This model was trained from scratch on an this dataset for FIM task. It achieves the following results on the evaluation set:

• Loss: 0.6155

Model description

Codellama-7b model trained for FIM on Java code dataset.

Intended uses & limitations

Bleh

Training and evaluation data

Dataset mentioned above

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: 0.0005
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 30
• training_steps: 2000

Training results

Training Loss	Epoch	Step	Validation Loss
0.6594	0.05	100	0.6927
0.6701	0.1	200	0.6784
0.6329	0.15	300	0.6690
0.6361	0.2	400	0.6629
0.5964	0.25	500	0.6545
0.6247	0.3	600	0.6461
0.6146	0.35	700	0.6407
0.5892	0.4	800	0.6364
0.5916	0.45	900	0.6308
0.6069	0.5	1000	0.6267
0.5804	0.55	1100	0.6242
0.5793	0.6	1200	0.6212
0.5836	0.65	1300	0.6195
0.5839	0.7	1400	0.6174
0.597	0.75	1500	0.6162
0.6042	0.8	1600	0.6158
0.5777	0.85	1700	0.6155
0.5683	0.9	1800	0.6155
0.5613	0.95	1900	0.6155
0.5597	1.0	2000	0.6155

Framework versions

• PEFT 0.5.0
• Transformers 4.34.0
• Pytorch 2.1.0+cu118
• Datasets 2.16.1
• Tokenizers 0.14.1