Views
No views yet
Llama-2-7b-hf on the henrichsen/alpaca_2k_test dataset.
Number of GPUs: 8x NVIDIA V100 GPUs
GPU Memory: 32GB each (SXM2 form factor)1base_model: meta-llama/Llama-2-7b-hf
2base_model_config: meta-llama/Llama-2-7b-hf
3model_type: LlamaForCausalLM
4tokenizer_type: LlamaTokenizer
5is_llama_derived_model: true
6
7load_in_8bit: true
8load_in_4bit: false
9strict: false
10
11datasets:
12 - path: mhenrichsen/alpaca_2k_test
13 type: alpaca
14dataset_prepared_path: last_run_prepared
15val_set_size: 0.01
16output_dir: ./lora-out
17
18sequence_len: 4096
19sample_packing: false
20pad_to_sequence_len: true
21
22adapter: lora
23lora_model_dir:
24lora_r: 32
25lora_alpha: 16
26lora_dropout: 0.05
27lora_target_linear: true
28lora_fan_in_fan_out:
29
30wandb_project:
31wandb_entity:
32wandb_watch:
33wandb_run_id:
34wandb_log_model:
35
36gradient_accumulation_steps: 4
37micro_batch_size: 2
38num_epochs: 3
39optimizer: adamw_bnb_8bit
40lr_scheduler: cosine
41learning_rate: 0.0002
42
43train_on_inputs: false
44group_by_length: false
45bf16: false
46fp16: true
47tf32: false
48
49gradient_checkpointing: true
50early_stopping_patience:
51resume_from_checkpoint:
52local_rank:
53logging_steps: 1
54xformers_attention: true
55flash_attention: false
56
57warmup_steps: 10
58eval_steps: 20
59save_steps:
60debug:
61deepspeed:
62weight_decay: 0.0
63fsdp:
64fsdp_config:
65special_tokens:
66 bos_token: "<s>"
67 eos_token: "</s>"
68 unk_token: "<unk>"The following `bitsandbytes` quantization config was used during training:
- quant_method: bitsandbytes
- load_in_8bit: True
- load_in_4bit: False
- 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: fp4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float32