[QEff Finetune]: Adding steps about how to fine tune on any custom dataset. by quic-swatia · Pull Request #381 · quic/efficient-transformers

quic-swatia · 2025-04-28T11:14:42Z

Added steps on how to create the custom_dataset.py to run fine-tuning through QEfficient pipeline on any custom dataset. Also, added a detailed template for the user which covers how to create custom_dataset.py
Added the argument 'context_length' in the existing APIs which helps run fine tuning with padding for custom dataset.
Made alpaca_dataset as the default dataset.
For DDP without sorting, shuffling was set to True. Made it False to sync it up with single SOC run and also to be able to use 'resume finetuning from between' feature.

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-meetkuma

Good work in listing down the detailed steps for custom dataset, Swati! Please check on the comments. :)

quic-meetkuma · 2025-05-12T09:44:36Z

+
+To run fine tuning for any user specific dataset, prepare the dataset using the following steps:
+
+    1) Create a  directory named 'dataset' inside efficient-transformers. 


double space between "a" and "directory"

quic-meetkuma · 2025-05-12T09:46:46Z

+    def apply_prompt_template():
+
+    def tokenize():
+


Add some comment as "Implement tokenization and prepare inputs for the training."

quic-meetkuma · 2025-05-12T09:49:04Z

+
+    # load dataset
+    # based on split, retrieve only the specific portion of the dataset (train or eval) either here or at the last
+


Add one more comment as "Define a prompt template"

It's already there. ( # define prompt)

quic-meetkuma · 2025-05-12T09:49:32Z

+    # based on split, retrieve only the specific portion of the dataset (train or eval) either here or at the last
+
+    def apply_prompt_template():
+


Add some comment as "Convert the raw input into format as per the template defined earlier."

quic-meetkuma · 2025-05-12T09:53:16Z

+    4) get_custom_dataset() should have following 4 parameters:  dataset_config, tokenizer, split, context_length. This function gets called twice through Qefficient/cloud/finetune.py with the name get_preprocessed_dataset. 
+    5) Inside get_custom_dataset(), dataset needs to prepared for fine tuning. So, the user needs to apply prompt and tokenize the dataset accordingly. Please refer the below template on how to define get_custom_dataset().
+    6) For examples, please refer python files present in efficient-transformers/QEfficient/finetune/dataset. In case of Samsum dataset, get_preprocessed_samsum() of efficient-transformers/QEfficient/finetune/dataset/samsum_dataset.py is called. 
+    7) In efficient-transformers/QEfficient/finetune/configs/dataset_config.py, for custom_dataset class, pass the appropriate value for train_split and test_split according to the dataset keys corresponding to train and test data points.


I think this is no longer needed after PR#289. We can directly pass --train_split and --test_split from the CLI.

Yes, correct! Thanks for pointing this. After PR 289, user can do this through both ways. Updated step #7 accordingly.

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-meetkuma · 2025-05-19T10:50:58Z

    use_autocast: bool = True
    val_batch_size: int = 1
-    dataset = "samsum_dataset"
+    dataset = "alpaca_dataset"


Good that you have added this change in this gerrit.

quic-meetkuma

Looks good to me.

quic-mamta · 2025-05-19T15:59:40Z

+
+To run fine tuning for any user specific dataset, prepare the dataset using the following steps:
+
+    1) Create a directory named 'dataset' inside efficient-transformers. 


Add the location "at root of the repo."

quic-mamta · 2025-05-19T16:00:42Z

+    1) Create a directory named 'dataset' inside efficient-transformers. 
+    2) Inside this directory, create a file named 'custom_dataset.py'. This is different than the custom_dataset.py present at efficient-transformers/QEfficient/finetune/dataset.
+    3) Inside the newly created efficient-transformers/dataset/custom_dataset.py, define a function named 'get_custom_dataset'. 
+    4) get_custom_dataset() should have following 4 parameters:  dataset_config, tokenizer, split, context_length. This function gets called twice through Qefficient/cloud/finetune.py with the name get_preprocessed_dataset. 


QEfficient not Qefficient

quic-mamta · 2025-05-19T16:01:32Z

+    3) Inside the newly created efficient-transformers/dataset/custom_dataset.py, define a function named 'get_custom_dataset'. 
+    4) get_custom_dataset() should have following 4 parameters:  dataset_config, tokenizer, split, context_length. This function gets called twice through Qefficient/cloud/finetune.py with the name get_preprocessed_dataset. 
+    5) Inside get_custom_dataset(), dataset needs to prepared for fine tuning. So, the user needs to apply prompt and tokenize the dataset accordingly. Please refer the below template on how to define get_custom_dataset().
+    6) For examples, please refer python files present in efficient-transformers/QEfficient/finetune/dataset. In case of Samsum dataset, get_preprocessed_samsum() of efficient-transformers/QEfficient/finetune/dataset/samsum_dataset.py is called. 


since default dataset is changed, we should mention alpaca here.

The steps I have mentioned matches with the format of samsum_dataset.py. It doesn't match with alpaca_dataset.py. Hence, I didn't change it.

too verbose. Make it simple pointed steps

Done. Added the detailed points in confluence and made them crisp in the PR.

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

anujgupt-github · 2025-05-21T03:26:32Z

+    3) Inside the newly created efficient-transformers/dataset/custom_dataset.py, define a function named 'get_custom_dataset'. 
+    4) get_custom_dataset() should have following 4 parameters:  dataset_config, tokenizer, split, context_length. This function gets called twice through Qefficient/cloud/finetune.py with the name get_preprocessed_dataset. 
+    5) Inside get_custom_dataset(), dataset needs to prepared for fine tuning. So, the user needs to apply prompt and tokenize the dataset accordingly. Please refer the below template on how to define get_custom_dataset().
+    6) For examples, please refer python files present in efficient-transformers/QEfficient/finetune/dataset. In case of Samsum dataset, get_preprocessed_samsum() of efficient-transformers/QEfficient/finetune/dataset/samsum_dataset.py is called. 


too verbose. Make it simple pointed steps

anujgupt-github · 2025-05-21T03:27:27Z

+    4) get_custom_dataset() should have following 4 parameters:  dataset_config, tokenizer, split, context_length. This function gets called twice through QEfficient/cloud/finetune.py with the name get_preprocessed_dataset. 
+    5) Inside get_custom_dataset(), dataset needs to prepared for fine tuning. So, the user needs to apply prompt and tokenize the dataset accordingly. Please refer the below template on how to define get_custom_dataset().
+    6) For examples, please refer python files present in efficient-transformers/QEfficient/finetune/dataset. In case of Samsum dataset, get_preprocessed_samsum() of efficient-transformers/QEfficient/finetune/dataset/samsum_dataset.py is called. 
+    7) In efficient-transformers/QEfficient/finetune/configs/dataset_config.py, for custom_dataset class, pass the appropriate value for train_split and test_split according to the dataset keys corresponding to train and test data points. As an alternative, these values can be passed as command line arguemnets as well with the finetune command. For example "--train_split train". 


Add hyperlinks to the relative paths annotated in the steps below

anujgupt-github · 2025-05-21T03:28:55Z

-```
+```
+
+## Fine-Tuning on custom dataset


You should include details on how we use gradient accumulation, how the dataset is shuffled, how activation checkpointing is enabled in separate sections.
In custom dataset, add a point that if any user wants to use these, refer xyz section

Added the section explaining how to use gradient accumulation and gradient checkpointing.
For single SOC, we are not doing any shuffling of data. By default, it is False.
For DDP, in case of sorting, shuffling is set to false.
For DDP without sorting, shuffling was set to True. Made it False to sync it up with single SOC run and also to be able to use 'resume finetuning from between' feature. this was not caught earlier because by default we run DDP with sorting only.
Hence, didn't add any information about shuffling in the doc.

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>

…taset. (quic#381) 1) Added steps on how to create the custom_dataset.py to run fine-tuning through QEfficient pipeline on any custom dataset. Also, added a detailed template for the user which covers how to create custom_dataset.py 2) Added the argument 'context_length' in the existing APIs which helps run fine tuning with padding for custom dataset. 3) Made alpaca_dataset as the default dataset. 4) For DDP without sorting, shuffling was set to True. Made it False to sync it up with single SOC run and also to be able to use 'resume finetuning from between' feature. --------- Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com> Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>

quic-swatia requested review from ochougul and quic-rishinr as code owners April 28, 2025 11:14

quic-swatia self-assigned this Apr 28, 2025

quic-swatia requested a review from vbaddi April 28, 2025 11:16

quic-rishinr marked this pull request as draft April 28, 2025 16:51

Adding steps about how to fine tune on any custom dataset.

e757ab8

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-swatia force-pushed the custom_dataset branch from 8485344 to e757ab8 Compare April 29, 2025 09:10

quic-mamta changed the title ~~Adding steps about how to fine tune on any custom dataset.~~ [QEff Finetune]: Adding steps about how to fine tune on any custom dataset. May 8, 2025

quic-swatia marked this pull request as ready for review May 9, 2025 10:39

quic-swatia requested a review from quic-amitraj as a code owner May 9, 2025 10:39

Update finetune.md

bae75d2

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-rishinr added the fine-tuning label May 12, 2025

quic-swatia requested a review from quic-mamta May 12, 2025 05:35

quic-meetkuma requested changes May 12, 2025

View reviewed changes

Adding alpaca_dataset as teh default dataset

05a385a

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-swatia force-pushed the custom_dataset branch from f46f537 to 05a385a Compare May 19, 2025 10:45

quic-meetkuma reviewed May 19, 2025

View reviewed changes

quic-meetkuma approved these changes May 19, 2025

View reviewed changes

quic-mamta reviewed May 19, 2025

View reviewed changes

Update finetune.md

155bb77

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-mamta approved these changes May 19, 2025

View reviewed changes

anujgupt-github reviewed May 21, 2025

View reviewed changes

quic-swatia force-pushed the custom_dataset branch from 43aab17 to 155bb77 Compare June 12, 2025 06:18

quic-swatia requested a review from quic-hemagnih as a code owner June 12, 2025 06:18

Merge branch 'main' into custom_dataset

6262009

Signed-off-by: Swati Allabadi <quic_sallabad@quicinc.com>

quic-swatia force-pushed the custom_dataset branch from ea0f63a to 1be1053 Compare June 12, 2025 17:44

Adding steps about how to fine tune on any custom dataset.

25a2ac5

Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>

quic-swatia force-pushed the custom_dataset branch from 1be1053 to 25a2ac5 Compare June 12, 2025 17:52

quic-swatia merged commit 64f4a04 into quic:main Jun 13, 2025


		To run fine tuning for any user specific dataset, prepare the dataset using the following steps:

		1) Create a directory named 'dataset' inside efficient-transformers.


		# load dataset
		# based on split, retrieve only the specific portion of the dataset (train or eval) either here or at the last

		# based on split, retrieve only the specific portion of the dataset (train or eval) either here or at the last

		def apply_prompt_template():

Uh oh!

Conversation

quic-swatia commented Apr 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

quic-meetkuma left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

quic-swatia May 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

quic-meetkuma left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

quic-swatia commented Apr 28, 2025 •

edited

Loading

quic-swatia May 14, 2025 •

edited

Loading