Commit
•
442de62
1
Parent(s):
83fe2e3
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,13 +1,91 @@
|
|
| 1 |
---
|
| 2 |
base_model: ai-forever/ruGPT-3.5-13B
|
| 3 |
library_name: peft
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
- IlyaGusev/ru_turbo_alpaca
|
| 9 |
- IlyaGusev/ru_turbo_alpaca_evol_instruct
|
| 10 |
- IlyaGusev/ru_turbo_saiga
|
| 11 |
- IlyaGusev/ru_sharegpt_cleaned
|
| 12 |
- IlyaGusev/oasst1_ru_main_branch
|
| 13 |
- lksy/ru_instruct_gpt4
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
base_model: ai-forever/ruGPT-3.5-13B
|
| 3 |
library_name: peft
|
| 4 |
+
license: mit
|
| 5 |
+
language:
|
| 6 |
+
- ru
|
| 7 |
+
tags:
|
| 8 |
+
- impruver
|
| 9 |
+
- russian
|
| 10 |
+
- function call
|
| 11 |
+
- lora
|
| 12 |
+
pipeline_tag: text-generation
|
| 13 |
+
datasets:
|
| 14 |
- IlyaGusev/ru_turbo_alpaca
|
| 15 |
- IlyaGusev/ru_turbo_alpaca_evol_instruct
|
| 16 |
- IlyaGusev/ru_turbo_saiga
|
| 17 |
- IlyaGusev/ru_sharegpt_cleaned
|
| 18 |
- IlyaGusev/oasst1_ru_main_branch
|
| 19 |
- lksy/ru_instruct_gpt4
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
# ruGPT-3.5-13B / Saiga2
|
| 23 |
+
|
| 24 |
+
LoRA адаптер для ruGPT3.5-13B обученный на коллекции датасетов Saiga.
|
| 25 |
+
|
| 26 |
+
Конфигурация: https://github.com/EvilFreelancer/impruver/blob/main/configs/ruGPT35_13B_lora.yml
|
| 27 |
+
|
| 28 |
+
Адаптер обучался на 1x RTX 4090, для этого потребовалось примерно 18.2Gb VRAM и заняло 16h 58m.
|
| 29 |
+
|
| 30 |
+
```yml
|
| 31 |
+
output_dir: ./models/ruGPT35_13B_lora
|
| 32 |
+
train_path: ./train.ruGPT35_13B.jsonl
|
| 33 |
+
val_path: ./val.ruGPT35_13B.jsonl
|
| 34 |
+
|
| 35 |
+
datasets:
|
| 36 |
+
- name: IlyaGusev/ru_turbo_alpaca
|
| 37 |
+
converter: impruver.instruction_to_messages
|
| 38 |
+
- name: IlyaGusev/ru_turbo_alpaca_evol_instruct
|
| 39 |
+
converter: impruver.instruction_to_messages
|
| 40 |
+
- name: IlyaGusev/ru_turbo_saiga
|
| 41 |
+
converter: impruver.dialog_to_messages
|
| 42 |
+
- name: IlyaGusev/ru_sharegpt_cleaned
|
| 43 |
+
converter: impruver.dialog_to_messages
|
| 44 |
+
- name: IlyaGusev/oasst1_ru_main_branch
|
| 45 |
+
converter: impruver.dialog_to_messages
|
| 46 |
+
- name: lksy/ru_instruct_gpt4
|
| 47 |
+
converter: impruver.converters.instruction_to_messages
|
| 48 |
+
|
| 49 |
+
model:
|
| 50 |
+
class: transformers.AutoModelForCausalLM
|
| 51 |
+
name: ai-forever/ruGPT-3.5-13B
|
| 52 |
+
load_in_4bit: true
|
| 53 |
+
load_in_8bit: false
|
| 54 |
+
dtype: bf16
|
| 55 |
+
|
| 56 |
+
lora:
|
| 57 |
+
r: 16
|
| 58 |
+
lora_alpha: 16
|
| 59 |
+
lora_dropout: 0.05
|
| 60 |
+
bias: none
|
| 61 |
+
target_modules: [ c_attn ]
|
| 62 |
+
task_type: CAUSAL_LM
|
| 63 |
+
|
| 64 |
+
tokenizer:
|
| 65 |
+
class: transformers.AutoTokenizer
|
| 66 |
+
name: ai-forever/ruGPT-3.5-13B
|
| 67 |
+
max_tokens_count: 1024
|
| 68 |
+
|
| 69 |
+
trainer:
|
| 70 |
+
eval_strategy: steps
|
| 71 |
+
save_strategy: steps
|
| 72 |
+
eval_steps: 100
|
| 73 |
+
save_steps: 100
|
| 74 |
+
per_device_train_batch_size: 1
|
| 75 |
+
per_device_eval_batch_size: 1
|
| 76 |
+
gradient_accumulation_steps: 128
|
| 77 |
+
logging_steps: 1
|
| 78 |
+
learning_rate: 0.0002
|
| 79 |
+
num_train_epochs: 2
|
| 80 |
+
lr_scheduler_type: cosine
|
| 81 |
+
warmup_steps: 16
|
| 82 |
+
optim: adamw_8bit
|
| 83 |
+
metric_for_best_model: eval_loss
|
| 84 |
+
load_best_model_at_end: true
|
| 85 |
+
save_total_limit: 2
|
| 86 |
+
seed: 42
|
| 87 |
+
remove_unused_columns: false
|
| 88 |
+
max_grad_norm: 1.0
|
| 89 |
+
weight_decay: 0.08
|
| 90 |
+
torch_compile: false
|
| 91 |
+
```
|