Commit
•
989c563
1
Parent(s):
d7957ee
Update README.md
Browse files
README.md
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
---
|
2 |
language:
|
3 |
- en
|
|
|
4 |
license: apache-2.0
|
5 |
tags:
|
6 |
- text-generation-inference
|
@@ -9,14 +10,81 @@ tags:
|
|
9 |
- llama
|
10 |
- trl
|
11 |
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
|
|
|
|
|
|
|
|
|
12 |
---
|
13 |
|
14 |
-
#
|
15 |
|
16 |
-
- **Developed by:** kirankunapuli
|
17 |
- **License:** apache-2.0
|
18 |
-
- **Finetuned from model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
21 |
|
22 |
-
[
|
|
|
|
|
|
1 |
---
|
2 |
language:
|
3 |
- en
|
4 |
+
- hi
|
5 |
license: apache-2.0
|
6 |
tags:
|
7 |
- text-generation-inference
|
|
|
10 |
- llama
|
11 |
- trl
|
12 |
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
|
13 |
+
datasets:
|
14 |
+
- yahma/alpaca-cleaned
|
15 |
+
- ravithejads/samvaad-hi-filtered
|
16 |
+
- HydraIndicLM/hindi_alpaca_dolly_67k
|
17 |
---
|
18 |
|
19 |
+
# TinyLlama-1.1B-Hinglish-LORA-v1.0 model
|
20 |
|
21 |
+
- **Developed by:** [Kiran Kunapuli](https://www.linkedin.com/in/kirankunapuli/)
|
22 |
- **License:** apache-2.0
|
23 |
+
- **Finetuned from model:** TinyLlama/TinyLlama-1.1B-Chat-v1.0
|
24 |
+
- **Model config:**
|
25 |
+
```python
|
26 |
+
model = FastLanguageModel.get_peft_model(
|
27 |
+
model,
|
28 |
+
r = 64,
|
29 |
+
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
|
30 |
+
"gate_proj", "up_proj", "down_proj",],
|
31 |
+
lora_alpha = 128,
|
32 |
+
lora_dropout = 0,
|
33 |
+
bias = "none",
|
34 |
+
use_gradient_checkpointing = True,
|
35 |
+
random_state = 42,
|
36 |
+
use_rslora = True,
|
37 |
+
loftq_config = None,
|
38 |
+
)
|
39 |
+
```
|
40 |
+
- **Training parameters:**
|
41 |
+
```python
|
42 |
+
trainer = SFTTrainer(
|
43 |
+
model = model,
|
44 |
+
tokenizer = tokenizer,
|
45 |
+
train_dataset = dataset,
|
46 |
+
dataset_text_field = "text",
|
47 |
+
max_seq_length = max_seq_length,
|
48 |
+
dataset_num_proc = 2,
|
49 |
+
packing = True,
|
50 |
+
args = TrainingArguments(
|
51 |
+
per_device_train_batch_size = 12,
|
52 |
+
gradient_accumulation_steps = 16,
|
53 |
+
warmup_ratio = 0.1,
|
54 |
+
num_train_epochs = 1,
|
55 |
+
learning_rate = 2e-4,
|
56 |
+
fp16 = not torch.cuda.is_bf16_supported(),
|
57 |
+
bf16 = torch.cuda.is_bf16_supported(),
|
58 |
+
logging_steps = 1,
|
59 |
+
optim = "paged_adamw_32bit",
|
60 |
+
weight_decay = 0.001,
|
61 |
+
lr_scheduler_type = "cosine",
|
62 |
+
seed = 42,
|
63 |
+
output_dir = "outputs",
|
64 |
+
report_to = "wandb",
|
65 |
+
),
|
66 |
+
)
|
67 |
+
```
|
68 |
+
- **Training details:**
|
69 |
+
```
|
70 |
+
==((====))== Unsloth - 2x faster free finetuning | Num GPUs = 1
|
71 |
+
\\ /| Num examples = 15,464 | Num Epochs = 1
|
72 |
+
O^O/ \_/ \ Batch size per device = 12 | Gradient Accumulation steps = 16
|
73 |
+
\ / Total batch size = 192 | Total steps = 80
|
74 |
+
"-____-" Number of trainable parameters = 50,462,720
|
75 |
+
|
76 |
+
GPU = NVIDIA GeForce RTX 3090. Max memory = 24.0 GB.
|
77 |
+
Total time taken for 1 epoch - 2h:35m:28s
|
78 |
+
9443.5288 seconds used for training.
|
79 |
+
157.39 minutes used for training.
|
80 |
+
Peak reserved memory = 17.641 GB.
|
81 |
+
Peak reserved memory for training = 15.344 GB.
|
82 |
+
Peak reserved memory % of max memory = 73.504 %.
|
83 |
+
Peak reserved memory for training % of max memory = 63.933 %.
|
84 |
+
```
|
85 |
|
86 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
87 |
|
88 |
+
**[NOTE]** TinyLlama's internal maximum sequence length is 2048. We use RoPE Scaling to extend it to 4096 with Unsloth!
|
89 |
+
|
90 |
+
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|