Update README.md
Browse files
README.md
CHANGED
@@ -1,110 +1,3 @@
|
|
1 |
---
|
2 |
-
library_name: peft
|
3 |
license: apache-2.0
|
4 |
-
|
5 |
-
- mhenrichsen/alpaca_2k_test
|
6 |
-
---
|
7 |
-
We fine tune base `Llama-2-7b-hf` on the `henrichsen/alpaca_2k_test` dataset using peft-LORA.
|
8 |
-
Find adapters at: https://huggingface.co/Tensoic/Llama-2-7B-alpaca-2k-test
|
9 |
-
|
10 |
-
Visit us at: https://tensoic.com
|
11 |
-
|
12 |
-
![image/png](https://cdn-uploads.huggingface.co/production/uploads/644bf6ef778ecbfb977e8e84/C0btqRI3eCz0kNYGQoa9k.png)
|
13 |
-
|
14 |
-
## Training Setup:
|
15 |
-
```
|
16 |
-
Number of GPUs: 8x NVIDIA V100 GPUs
|
17 |
-
GPU Memory: 32GB each (SXM2 form factor)
|
18 |
-
```
|
19 |
-
## Training Configuration:
|
20 |
-
|
21 |
-
```yaml
|
22 |
-
base_model: meta-llama/Llama-2-7b-hf
|
23 |
-
base_model_config: meta-llama/Llama-2-7b-hf
|
24 |
-
model_type: LlamaForCausalLM
|
25 |
-
tokenizer_type: LlamaTokenizer
|
26 |
-
is_llama_derived_model: true
|
27 |
-
|
28 |
-
load_in_8bit: true
|
29 |
-
load_in_4bit: false
|
30 |
-
strict: false
|
31 |
-
|
32 |
-
datasets:
|
33 |
-
- path: mhenrichsen/alpaca_2k_test
|
34 |
-
type: alpaca
|
35 |
-
dataset_prepared_path: last_run_prepared
|
36 |
-
val_set_size: 0.01
|
37 |
-
output_dir: ./lora-out
|
38 |
-
|
39 |
-
sequence_len: 4096
|
40 |
-
sample_packing: false
|
41 |
-
pad_to_sequence_len: true
|
42 |
-
|
43 |
-
adapter: lora
|
44 |
-
lora_model_dir:
|
45 |
-
lora_r: 32
|
46 |
-
lora_alpha: 16
|
47 |
-
lora_dropout: 0.05
|
48 |
-
lora_target_linear: true
|
49 |
-
lora_fan_in_fan_out:
|
50 |
-
|
51 |
-
wandb_project:
|
52 |
-
wandb_entity:
|
53 |
-
wandb_watch:
|
54 |
-
wandb_run_id:
|
55 |
-
wandb_log_model:
|
56 |
-
|
57 |
-
gradient_accumulation_steps: 4
|
58 |
-
micro_batch_size: 2
|
59 |
-
num_epochs: 3
|
60 |
-
optimizer: adamw_bnb_8bit
|
61 |
-
lr_scheduler: cosine
|
62 |
-
learning_rate: 0.0002
|
63 |
-
|
64 |
-
train_on_inputs: false
|
65 |
-
group_by_length: false
|
66 |
-
bf16: false
|
67 |
-
fp16: true
|
68 |
-
tf32: false
|
69 |
-
|
70 |
-
gradient_checkpointing: true
|
71 |
-
early_stopping_patience:
|
72 |
-
resume_from_checkpoint:
|
73 |
-
local_rank:
|
74 |
-
logging_steps: 1
|
75 |
-
xformers_attention: true
|
76 |
-
flash_attention: false
|
77 |
-
|
78 |
-
warmup_steps: 10
|
79 |
-
eval_steps: 20
|
80 |
-
save_steps:
|
81 |
-
debug:
|
82 |
-
deepspeed:
|
83 |
-
weight_decay: 0.0
|
84 |
-
fsdp:
|
85 |
-
fsdp_config:
|
86 |
-
special_tokens:
|
87 |
-
bos_token: "<s>"
|
88 |
-
eos_token: "</s>"
|
89 |
-
unk_token: "<unk>"
|
90 |
-
```
|
91 |
-
```
|
92 |
-
The following `bitsandbytes` quantization config was used during training:
|
93 |
-
- quant_method: bitsandbytes
|
94 |
-
- load_in_8bit: True
|
95 |
-
- load_in_4bit: False
|
96 |
-
- llm_int8_threshold: 6.0
|
97 |
-
- llm_int8_skip_modules: None
|
98 |
-
- llm_int8_enable_fp32_cpu_offload: False
|
99 |
-
- llm_int8_has_fp16_weight: False
|
100 |
-
- bnb_4bit_quant_type: fp4
|
101 |
-
- bnb_4bit_use_double_quant: False
|
102 |
-
- bnb_4bit_compute_dtype: float32
|
103 |
-
|
104 |
-
```
|
105 |
-
|
106 |
-
|
107 |
-
### Framework versions
|
108 |
-
|
109 |
-
|
110 |
-
- PEFT 0.6.0.dev0
|
|
|
1 |
---
|
|
|
2 |
license: apache-2.0
|
3 |
+
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|