Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base model = beomi-Llama-3-Open-Ko-8B-Instruct-preview
|
2 |
+
base model = hansoldeco-beomi-Llama-3-Open-Ko-8B-Instruct-preview (Trained via Axolotl)
|
3 |
+
|
4 |
+
dora_train config
|
5 |
+
(from fsdp_qlora repo)
|
6 |
+
|
7 |
+
'''
|
8 |
+
export CUDA_VISIBLE_DEVICES=0,1
|
9 |
+
python train.py \
|
10 |
+
--train_type bnb_dora \
|
11 |
+
--model_name sosoai/hansoldeco-beomi-Llama-3-Open-Ko-8B-Instruct-preview \
|
12 |
+
--dataset orca_math \
|
13 |
+
--dataset_samples 193789 \
|
14 |
+
--batch_size 4 \
|
15 |
+
--context_length 8192 \
|
16 |
+
--gradient_accumulation_steps 2 \
|
17 |
+
--sharding_strategy full_shard \
|
18 |
+
--use_gradient_checkpointing true \
|
19 |
+
--reentrant_checkpointing true \
|
20 |
+
--use_cpu_offload false \
|
21 |
+
--use_activation_cpu_offload false \
|
22 |
+
--log_to wandb \
|
23 |
+
--project_name "sosoai-fsdp-quantized-ft-exps" \
|
24 |
+
--save_model true \
|
25 |
+
--output_dir models/llama-8b-orca-math-10k-bnb-QDoRA
|
26 |
+
'''
|
27 |
+
|
28 |
+
Dataset = hansoldeco domain own dataset (Non open)
|
29 |
+
Dataset = kuotient/orca-math-word-problems-193k-korean
|
30 |
+
|