aligntune 0.2.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- aligntune-0.2.0/LICENSE +21 -0
- aligntune-0.2.0/PKG-INFO +306 -0
- aligntune-0.2.0/README.md +230 -0
- aligntune-0.2.0/pyproject.toml +361 -0
- aligntune-0.2.0/setup.cfg +4 -0
- aligntune-0.2.0/setup.py +10 -0
- aligntune-0.2.0/src/aligntune/__init__.py +368 -0
- aligntune-0.2.0/src/aligntune/__main__.py +11 -0
- aligntune-0.2.0/src/aligntune/_fallbacks.py +132 -0
- aligntune-0.2.0/src/aligntune/_imports.py +598 -0
- aligntune-0.2.0/src/aligntune/backends/__init__.py +64 -0
- aligntune-0.2.0/src/aligntune/backends/trl/__init__.py +55 -0
- aligntune-0.2.0/src/aligntune/backends/trl/eval/__init__.py +0 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rewards/__init__.py +27 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rewards/training.py +838 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/__init__.py +53 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/counterfact_grpo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/counterfact_grpo/counterfact_grpo.py +1283 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/counterfact_grpo/custom_trainer.py +2101 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dapo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dapo/dapo.py +1003 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dpo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dpo/dpo.py +668 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dr_grpo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/dr_grpo/drgrpo.py +939 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/gbmpo/__init__.py +3 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/gbmpo/gbmpo.py +1116 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/gbmpo/gbmpo_trainer.py +1826 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/grpo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/grpo/grpo.py +1105 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/gspo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/gspo/gspo.py +1381 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/es_utils.py +446 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/logging_config.py +70 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/meta_es_trainer.py +172 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/meta_es_trainer_utils.py +673 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/neural_mirror_grpo.py +726 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/train_nmdrgrpo_code.py +1523 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/train_nmdrgrpo_es_math.py +635 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/vllm_evaluator.py +674 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/meta_es/vllm_evaluator_math.py +489 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/neural_mirror_grpo/NMGrpo.py +941 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/neural_mirror_grpo/neural_mirror_grpo.py +726 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/pace/__init__.py +49 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/pace/baseline.py +549 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/pace/curriculum.py +467 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/pace/pace.py +981 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/ppo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/rl/ppo/ppo.py +1235 -0
- aligntune-0.2.0/src/aligntune/backends/trl/sft/Classification_trainer.py +787 -0
- aligntune-0.2.0/src/aligntune/backends/trl/sft/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/trl/sft/sft.py +1659 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/__init__.py +44 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/eval/__init__.py +0 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/__init__.py +36 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/bolt/__init__.py +18 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/bolt/bolt.py +695 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/counterfact_grpo/__init__.py +11 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/counterfact_grpo/counterfact_grpo.py +960 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dapo/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dapo/dapo.py +1060 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dpo/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dpo/dpo.py +712 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dr_grpo/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/dr_grpo/drgrpo.py +963 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/gbmpo/__init__.py +3 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/gbmpo/gbmpo.py +1216 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/gbmpo/gbmpo_trainer.py +1826 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/grpo/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/grpo/grpo.py +1108 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/gspo/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/gspo/gspo.py +1279 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/neural_mirror_grpo/NMGrpo.py +1140 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/neural_mirror_grpo/neural_mirror_grpo.py +726 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/ppo/__init__.py +76 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/ppo/ppo.py +2345 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/rl/ppo/unsloth_patches.py +355 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/sft/__init__.py +9 -0
- aligntune-0.2.0/src/aligntune/backends/unsloth/sft/sft.py +1203 -0
- aligntune-0.2.0/src/aligntune/cli/__init__.py +95 -0
- aligntune-0.2.0/src/aligntune/cli/__main__.py +10 -0
- aligntune-0.2.0/src/aligntune/cli/arg_parser.py +442 -0
- aligntune-0.2.0/src/aligntune/cli/config_builders.py +469 -0
- aligntune-0.2.0/src/aligntune/cli/diagnose.py +367 -0
- aligntune-0.2.0/src/aligntune/cli/finetune.py +508 -0
- aligntune-0.2.0/src/aligntune/cli/recipes.py +320 -0
- aligntune-0.2.0/src/aligntune/cli/unified-old.py +1135 -0
- aligntune-0.2.0/src/aligntune/cli/unified.py +1679 -0
- aligntune-0.2.0/src/aligntune/cli/validate.py +511 -0
- aligntune-0.2.0/src/aligntune/core/backend_factory.py +1320 -0
- aligntune-0.2.0/src/aligntune/core/callbacks/__init__.py +7 -0
- aligntune-0.2.0/src/aligntune/core/callbacks/base.py +165 -0
- aligntune-0.2.0/src/aligntune/core/dataset_adapters.py +361 -0
- aligntune-0.2.0/src/aligntune/core/optimization.py +424 -0
- aligntune-0.2.0/src/aligntune/core/precision_handler.py +207 -0
- aligntune-0.2.0/src/aligntune/core/rl/__init__.py +65 -0
- aligntune-0.2.0/src/aligntune/core/rl/caching.py +324 -0
- aligntune-0.2.0/src/aligntune/core/rl/config.py +729 -0
- aligntune-0.2.0/src/aligntune/core/rl/config_loader.py +307 -0
- aligntune-0.2.0/src/aligntune/core/rl/distributed.py +319 -0
- aligntune-0.2.0/src/aligntune/core/rl/evaluator.py +1053 -0
- aligntune-0.2.0/src/aligntune/core/rl/function_based_reward_model.py +276 -0
- aligntune-0.2.0/src/aligntune/core/rl/logging.py +327 -0
- aligntune-0.2.0/src/aligntune/core/rl/models.py +371 -0
- aligntune-0.2.0/src/aligntune/core/rl/registries.py +631 -0
- aligntune-0.2.0/src/aligntune/core/rl/reward_model_wrapper.py +212 -0
- aligntune-0.2.0/src/aligntune/core/rl/rollout.py +412 -0
- aligntune-0.2.0/src/aligntune/core/rl/sample_logger.py +135 -0
- aligntune-0.2.0/src/aligntune/core/rl/trainer_base.py +920 -0
- aligntune-0.2.0/src/aligntune/core/rl/trainer_factory.py +152 -0
- aligntune-0.2.0/src/aligntune/core/sft/__init__.py +44 -0
- aligntune-0.2.0/src/aligntune/core/sft/config.py +762 -0
- aligntune-0.2.0/src/aligntune/core/sft/config_loader.py +174 -0
- aligntune-0.2.0/src/aligntune/core/sft/evaluator.py +1017 -0
- aligntune-0.2.0/src/aligntune/core/sft/logging.py +117 -0
- aligntune-0.2.0/src/aligntune/core/sft/trainer_base.py +601 -0
- aligntune-0.2.0/src/aligntune/core/sft/trainer_factory.py +78 -0
- aligntune-0.2.0/src/aligntune/data/__init__.py +0 -0
- aligntune-0.2.0/src/aligntune/data/full_requirements_test.py +201 -0
- aligntune-0.2.0/src/aligntune/data/loaders/__init__.py +0 -0
- aligntune-0.2.0/src/aligntune/data/loaders/base.py +7 -0
- aligntune-0.2.0/src/aligntune/data/loaders/csv_loader.py +20 -0
- aligntune-0.2.0/src/aligntune/data/loaders/directory_loader.py +27 -0
- aligntune-0.2.0/src/aligntune/data/loaders/hf_loader.py +45 -0
- aligntune-0.2.0/src/aligntune/data/loaders/json_loader.py +12 -0
- aligntune-0.2.0/src/aligntune/data/loaders/parquet_loader.py +9 -0
- aligntune-0.2.0/src/aligntune/data/loaders/resolver.py +33 -0
- aligntune-0.2.0/src/aligntune/data/manager.py +214 -0
- aligntune-0.2.0/src/aligntune/data/processors.py +675 -0
- aligntune-0.2.0/src/aligntune/data/schemas.py +90 -0
- aligntune-0.2.0/src/aligntune/eval/__init__.py +113 -0
- aligntune-0.2.0/src/aligntune/eval/caching.py +40 -0
- aligntune-0.2.0/src/aligntune/eval/cli.py +248 -0
- aligntune-0.2.0/src/aligntune/eval/core.py +419 -0
- aligntune-0.2.0/src/aligntune/eval/evaluator.py +492 -0
- aligntune-0.2.0/src/aligntune/eval/full_evaluation_test.py +228 -0
- aligntune-0.2.0/src/aligntune/eval/lm_eval_integration.py +443 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/__init__.py +22 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/base.py +45 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/code.py +249 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/dpo.py +372 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/generic.py +52 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/math.py +139 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/rl.py +57 -0
- aligntune-0.2.0/src/aligntune/eval/metrics/text.py +128 -0
- aligntune-0.2.0/src/aligntune/eval/registry.py +56 -0
- aligntune-0.2.0/src/aligntune/eval/rl_evaluator.py +573 -0
- aligntune-0.2.0/src/aligntune/eval/runner.py +751 -0
- aligntune-0.2.0/src/aligntune/eval/safe_executor.py +509 -0
- aligntune-0.2.0/src/aligntune/main.py +1024 -0
- aligntune-0.2.0/src/aligntune/py.typed +0 -0
- aligntune-0.2.0/src/aligntune/recipes/__init__.py +538 -0
- aligntune-0.2.0/src/aligntune/recipes/config.py +22 -0
- aligntune-0.2.0/src/aligntune/rewards/__init__.py +69 -0
- aligntune-0.2.0/src/aligntune/rewards/core.py +3161 -0
- aligntune-0.2.0/src/aligntune/rewards/factory.py +162 -0
- aligntune-0.2.0/src/aligntune/rewards/registry.py +528 -0
- aligntune-0.2.0/src/aligntune/rewards/training.py +804 -0
- aligntune-0.2.0/src/aligntune/rl/__init__.py +10 -0
- aligntune-0.2.0/src/aligntune/rl/core/__init__.py +10 -0
- aligntune-0.2.0/src/aligntune/scripts/__init__.py +3 -0
- aligntune-0.2.0/src/aligntune/scripts/precompute_baseline.py +1937 -0
- aligntune-0.2.0/src/aligntune/sft/__init__.py +10 -0
- aligntune-0.2.0/src/aligntune/sft/core/__init__.py +10 -0
- aligntune-0.2.0/src/aligntune/utils/__init__.py +132 -0
- aligntune-0.2.0/src/aligntune/utils/auth.py +239 -0
- aligntune-0.2.0/src/aligntune/utils/checkpointing.py +439 -0
- aligntune-0.2.0/src/aligntune/utils/colored_logging.py +259 -0
- aligntune-0.2.0/src/aligntune/utils/config_extractor.py +346 -0
- aligntune-0.2.0/src/aligntune/utils/config_utils.py +501 -0
- aligntune-0.2.0/src/aligntune/utils/device.py +425 -0
- aligntune-0.2.0/src/aligntune/utils/diagnostics.py +454 -0
- aligntune-0.2.0/src/aligntune/utils/environment.py +278 -0
- aligntune-0.2.0/src/aligntune/utils/errors.py +429 -0
- aligntune-0.2.0/src/aligntune/utils/inference_utils.py +40 -0
- aligntune-0.2.0/src/aligntune/utils/logging.py +301 -0
- aligntune-0.2.0/src/aligntune/utils/math_grading.py +758 -0
- aligntune-0.2.0/src/aligntune/utils/model_loader.py +577 -0
- aligntune-0.2.0/src/aligntune/utils/validation.py +514 -0
- aligntune-0.2.0/src/aligntune.egg-info/PKG-INFO +306 -0
- aligntune-0.2.0/src/aligntune.egg-info/SOURCES.txt +183 -0
- aligntune-0.2.0/src/aligntune.egg-info/dependency_links.txt +1 -0
- aligntune-0.2.0/src/aligntune.egg-info/entry_points.txt +25 -0
- aligntune-0.2.0/src/aligntune.egg-info/requires.txt +47 -0
- aligntune-0.2.0/src/aligntune.egg-info/top_level.txt +1 -0
aligntune-0.2.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2025 Lexsi.ai
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
aligntune-0.2.0/PKG-INFO
ADDED
|
@@ -0,0 +1,306 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: aligntune
|
|
3
|
+
Version: 0.2.0
|
|
4
|
+
Summary: AlignTune: Multi-backend alignment and fine-tuning library (formerly FinetuneHub). Features TRL and Unsloth backends with complete RL coverage (DPO, PPO, GRPO, BOLT), 27+ reward functions, production-ready evaluation system, and unified configuration interface for LLM alignment.
|
|
5
|
+
Author-email: Chirag Chawla <chirag@lexsi.ai>, Zera Lyngkhoi <zera@lexsi.ai>, Pratinav Seth <pratinav@lexsi.ai>, Utsav Avaiya <utsav@lexsi.ai>, Soham Bhattacharjee <soham@lexsi.ai>, Mykola Khandoga <mykola@lexsi.ai>, Rui Yuan <rui@lexsi.ai>, Vinay Kumar Sankarapu <vinay@lexsi.ai>
|
|
6
|
+
License: AlignTune Source Available License (ASAL) v1.0
|
|
7
|
+
Project-URL: Homepage, https://github.com/Lexsi-Labs/aligntune
|
|
8
|
+
Project-URL: Repository, https://github.com/Lexsi-Labs/aligntune
|
|
9
|
+
Project-URL: Documentation, https://aligntune.readthedocs.io
|
|
10
|
+
Project-URL: Bug Tracker, https://github.com/Lexsi-Labs/aligntune/issues
|
|
11
|
+
Project-URL: Changelog, https://github.com/Lexsi-Labs/aligntune/blob/main/CHANGELOG.md
|
|
12
|
+
Project-URL: Discussions, https://github.com/Lexsi-Labs/aligntune/discussions
|
|
13
|
+
Keywords: alignment,rlhf,fine-tuning,llm,machine learning,SFT,RL,DPO,GRPO,PPO,GSPO,BOLT,transformers,trl,unsloth,multi-backend,reward-functions,evaluation,lm-eval,wandb,tensorboard,checkpointing,device-management,unified-config,cli
|
|
14
|
+
Classifier: Development Status :: 4 - Beta
|
|
15
|
+
Classifier: Intended Audience :: Developers
|
|
16
|
+
Classifier: Intended Audience :: Science/Research
|
|
17
|
+
Classifier: License :: Other/Proprietary License
|
|
18
|
+
Classifier: Programming Language :: Python :: 3
|
|
19
|
+
Classifier: Programming Language :: Python :: 3.8
|
|
20
|
+
Classifier: Programming Language :: Python :: 3.9
|
|
21
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
22
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
23
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
24
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
25
|
+
Classifier: Topic :: Software Development :: Libraries :: Python Modules
|
|
26
|
+
Classifier: Operating System :: OS Independent
|
|
27
|
+
Requires-Python: >=3.8
|
|
28
|
+
Description-Content-Type: text/markdown
|
|
29
|
+
License-File: LICENSE
|
|
30
|
+
Requires-Dist: transformers>=4.35.0
|
|
31
|
+
Requires-Dist: datasets>=2.14.0
|
|
32
|
+
Requires-Dist: accelerate>=0.24.0
|
|
33
|
+
Requires-Dist: peft>=0.6.0
|
|
34
|
+
Requires-Dist: bitsandbytes>=0.41.0
|
|
35
|
+
Requires-Dist: numpy>=1.21.0
|
|
36
|
+
Requires-Dist: pyyaml>=6.0
|
|
37
|
+
Requires-Dist: huggingface-hub>=0.17.0
|
|
38
|
+
Requires-Dist: tqdm>=4.64.0
|
|
39
|
+
Requires-Dist: psutil>=5.9.0
|
|
40
|
+
Requires-Dist: pandas>=1.3.0
|
|
41
|
+
Requires-Dist: scikit-learn>=1.0.0
|
|
42
|
+
Requires-Dist: click>=8.0.0
|
|
43
|
+
Requires-Dist: rich>=12.0.0
|
|
44
|
+
Requires-Dist: typer>=0.7.0
|
|
45
|
+
Requires-Dist: trl==0.23.0
|
|
46
|
+
Requires-Dist: unsloth==2025.11.1
|
|
47
|
+
Requires-Dist: unsloth_zoo==2025.11.2
|
|
48
|
+
Requires-Dist: wandb>=0.15.0
|
|
49
|
+
Requires-Dist: tensorboard>=2.12.0
|
|
50
|
+
Requires-Dist: matplotlib>=3.5.0
|
|
51
|
+
Requires-Dist: seaborn>=0.11.0
|
|
52
|
+
Requires-Dist: evaluate>=0.4.0
|
|
53
|
+
Requires-Dist: rouge-score>=0.1.2
|
|
54
|
+
Requires-Dist: sacrebleu>=2.3.0
|
|
55
|
+
Requires-Dist: lm-eval>=0.4.0
|
|
56
|
+
Requires-Dist: nltk>=3.8.0
|
|
57
|
+
Requires-Dist: nvitop>=1.1.0; platform_system != "Darwin"
|
|
58
|
+
Requires-Dist: nvidia-ml-py>=11.495.46; platform_system != "Darwin"
|
|
59
|
+
Requires-Dist: gpustat>=1.1.1
|
|
60
|
+
Requires-Dist: pytest>=7.0.0
|
|
61
|
+
Requires-Dist: pytest-cov>=4.0.0
|
|
62
|
+
Requires-Dist: pytest-xdist>=3.0.0
|
|
63
|
+
Requires-Dist: pytest-mock>=3.10.0
|
|
64
|
+
Requires-Dist: pytest-timeout>=2.1.0
|
|
65
|
+
Requires-Dist: hypothesis>=6.70.0
|
|
66
|
+
Requires-Dist: black>=23.0.0
|
|
67
|
+
Requires-Dist: isort>=5.12.0
|
|
68
|
+
Requires-Dist: flake8>=6.0.0
|
|
69
|
+
Requires-Dist: mypy>=1.5.0
|
|
70
|
+
Requires-Dist: pre-commit>=3.0.0
|
|
71
|
+
Requires-Dist: sphinx>=5.0.0
|
|
72
|
+
Requires-Dist: sphinx-rtd-theme>=1.2.0
|
|
73
|
+
Requires-Dist: myst-parser>=0.18.0
|
|
74
|
+
Requires-Dist: sphinx-autodoc-typehints>=1.19.0
|
|
75
|
+
Dynamic: license-file
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
<p align="center">
|
|
79
|
+
<img src="docs/assets/aligntune-banner.png" alt="AlignTune Banner" width="1000px"/>
|
|
80
|
+
</p>
|
|
81
|
+
|
|
82
|
+
<div align="center">
|
|
83
|
+
<a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.8+-blue.svg"/></a>
|
|
84
|
+
<a href="https://github.com/Lexsi-Labs/aligntune/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg"/></a>
|
|
85
|
+
<a href="https://badge.fury.io/py/aligntune"><img src="https://badge.fury.io/py/aligntune.svg"/></a>
|
|
86
|
+
</div>
|
|
87
|
+
|
|
88
|
+
---
|
|
89
|
+
|
|
90
|
+
**AlignTune** is a production-ready fine-tuning library designed to simplify training and fine-tuning of Large Language Models (LLMs) with both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) methods. It provides a high-level, unified API that abstracts away the complexities of backend selection, algorithm configuration, and training loops, letting you focus on delivering results.
|
|
91
|
+
|
|
92
|
+
## Core Features
|
|
93
|
+
|
|
94
|
+
**Multi-Backend Architecture**: Choose between TRL (reliable, battle-tested) and Unsloth (faster) backends with intelligent auto-selection.
|
|
95
|
+
|
|
96
|
+
**Complete RLHF Coverage**: 12+ RL algorithms including DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, and PACE.
|
|
97
|
+
|
|
98
|
+
**Production-Ready**: No mock code, comprehensive error handling, extensive testing, and robust validation.
|
|
99
|
+
|
|
100
|
+
## Quick Start
|
|
101
|
+
|
|
102
|
+
### Supervised Fine-Tuning (SFT)
|
|
103
|
+
|
|
104
|
+
```python
|
|
105
|
+
from aligntune.core.backend_factory import create_sft_trainer
|
|
106
|
+
|
|
107
|
+
# Create and train SFT model
|
|
108
|
+
trainer = create_sft_trainer(
|
|
109
|
+
model_name="microsoft/DialoGPT-small",
|
|
110
|
+
dataset_name="tatsu-lab/alpaca",
|
|
111
|
+
backend="trl",
|
|
112
|
+
num_epochs=3,
|
|
113
|
+
max_steps = -1,
|
|
114
|
+
batch_size=4,
|
|
115
|
+
learning_rate=5e-5
|
|
116
|
+
)
|
|
117
|
+
|
|
118
|
+
# Train the model
|
|
119
|
+
trainer.train()
|
|
120
|
+
|
|
121
|
+
# Evaluate
|
|
122
|
+
metrics = trainer.evaluate()
|
|
123
|
+
print(metrics)
|
|
124
|
+
```
|
|
125
|
+
|
|
126
|
+
### Reinforcement Learning (DPO)
|
|
127
|
+
|
|
128
|
+
```python
|
|
129
|
+
from aligntune.core.backend_factory import create_rl_trainer
|
|
130
|
+
|
|
131
|
+
# Create and train DPO model
|
|
132
|
+
trainer = create_rl_trainer(
|
|
133
|
+
model_name="Qwen/Qwen3-0.6B",
|
|
134
|
+
dataset_name="Anthropic/hh-rlhf",
|
|
135
|
+
algorithm="dpo",
|
|
136
|
+
backend="trl",
|
|
137
|
+
num_epochs=1,
|
|
138
|
+
batch_size=4,
|
|
139
|
+
learning_rate=5e-5
|
|
140
|
+
)
|
|
141
|
+
|
|
142
|
+
# Train the model
|
|
143
|
+
trainer.train()
|
|
144
|
+
```
|
|
145
|
+
|
|
146
|
+
## Supported Algorithms
|
|
147
|
+
|
|
148
|
+
| Algorithm | TRL Backend | Unsloth Backend | Description |
|
|
149
|
+
|-----------|-------------|-----------------|-------------|
|
|
150
|
+
| **SFT** | Yes | Yes | Supervised Fine-Tuning |
|
|
151
|
+
| **DPO** | Yes | Yes | Direct Preference Optimization |
|
|
152
|
+
| **PPO** | Yes | Yes | Proximal Policy Optimization |
|
|
153
|
+
| **GRPO** | Yes | Yes | Group Relative Policy Optimization |
|
|
154
|
+
| **GSPO** | Yes | Yes | Group Sequential Policy Optimization |
|
|
155
|
+
| **DAPO** | Yes | Yes | Decouple Clip and Dynamic sAmpling Policy Optimization |
|
|
156
|
+
| **Dr. GRPO** | Yes | Yes | GRPO Done Right (unbiased variant) |
|
|
157
|
+
| **GBMPO** | Yes | No | Group-Based Mirror Policy Optimization |
|
|
158
|
+
| **Counterfactual GRPO** | Yes | Yes | Counterfactual GRPO variant |
|
|
159
|
+
| **PACE** | Yes | Yes | Baseline-Optimized Learning Technique |
|
|
160
|
+
|
|
161
|
+
## Installation
|
|
162
|
+
|
|
163
|
+
```bash
|
|
164
|
+
# Or install from source
|
|
165
|
+
git clone https://github.com/Lexsi-Labs/aligntune.git
|
|
166
|
+
cd aligntune
|
|
167
|
+
pip install -e .
|
|
168
|
+
```
|
|
169
|
+
|
|
170
|
+
### Requirements
|
|
171
|
+
|
|
172
|
+
- Python 3.8+
|
|
173
|
+
- PyTorch 2.0+
|
|
174
|
+
- CUDA-compatible GPU (recommended for faster training)
|
|
175
|
+
|
|
176
|
+
## Demo Notebooks
|
|
177
|
+
|
|
178
|
+
Interactive Colab notebooks demonstrating various AlignTune workflows:
|
|
179
|
+
Here are the organized tables containing the Colab links, models, and datasets provided in your text.
|
|
180
|
+
|
|
181
|
+
### Supervised Fine-Tuning (SFT)
|
|
182
|
+
|
|
183
|
+
| Backend| Model | Dataset | Link |
|
|
184
|
+
| --- | --- | --- | --- |
|
|
185
|
+
| TRL | **Qwen/Qwen3-4B-Instruct-2507** | sohamb37lexsi/bitext-wealth-management-llm-chatbot-splits | [](https://colab.research.google.com/drive/1okAnfMlkch-G5Dy2dmj_dfR07rtxmYoq?usp=sharing) |
|
|
186
|
+
| TRL | **Qwen3-4B-Instruct** | sohamb37lexsi/bitext-retail-banking-llm-chatbot-splits | [](https://colab.research.google.com/drive/1JnFvEWZ6PDqrDAyWznCdOA9oNn8aqN2v?usp=sharing) |
|
|
187
|
+
| Unsloth| **Qwen/Qwen2.5-0.5B-Instruct** | bebechien/MobileGameNPC | [](https://colab.research.google.com/drive/15R3JOrzAUuMagCamDHsqe0wNyb44rw2j?usp=sharing) |
|
|
188
|
+
| TRL | **google/txgemma-2b-predict** | trialbench_adverse-event-rate-prediction | [](https://colab.research.google.com/drive/1r91SS7Lb5LjkjzWXp7uzkpR3vlWZ6gwD?usp=sharing) |
|
|
189
|
+
| Unsloth| **Qwen/Qwen2.5-0.5B-Instruct** | bebechien/MobileGameNP |[](https://colab.research.google.com/drive/11DrRbG30MUCegZTDXwR9nxEdvKaLfWCb?usp=sharing) |
|
|
190
|
+
|
|
191
|
+
### Reinforcement Learning (RL)
|
|
192
|
+
|
|
193
|
+
| Backend| Algorithm | Model | Dataset | Link |
|
|
194
|
+
| --- | --- | --- | --- | --- |
|
|
195
|
+
| Unsloth| **DPO** | microsoft/phi-2 | argilla/distilabel-intel-orca-dpo-pairs | [](https://colab.research.google.com/drive/1aKdQzT7KGs0PEr7pO9rOXFkwQK6LzyRZ#scrollTo=dOnIJIAMxP9J) |
|
|
196
|
+
| TRL | **DPO** | google/gemma-2-2b-it | Anthropic/hh-rlhf | [](https://colab.research.google.com/drive/1WQhek61Z0v1zHFWT4mQwVX7n10ln-j3d?usp=sharing) |
|
|
197
|
+
| TRL | **DPO** | sohamb37lexsi/wealth_management_Qwen3-4B-Instruct-2507 | sohamb37lexsi/bitext_wealth_management_preference_data | [](https://colab.research.google.com/drive/1LR3KWcjQAFjSHf0MR9271ShchjPUVSpS?usp=sharing) |
|
|
198
|
+
| Unsloth| **PPO** | Qwen/Qwen2.5-0.5B-Instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1G4fdtO4DhBNwjOYhuaGmvIrBU_lc159H?usp=sharing) |
|
|
199
|
+
| TRL | **PPO** | EleutherAI/pythia-1.4b | CarperAI/openai_summarize_tldr | [](https://colab.research.google.com/drive/1wlbSvQAJv8ZFM2qGD4XlUOKPzZHXstQo?usp=sharing) |
|
|
200
|
+
| TRL | **GRPO** (Coding) | Qwen/Qwen3-4B | google-research-datasets/mbpp | [](https://colab.research.google.com/drive/13HYZ-EkLC3-6wxJG_1NyWXPEuN9bIeM4?usp=sharing) |
|
|
201
|
+
| Unsloth| **GRPO** (Math) | meta-llama/Llama-3.2-3B-Instruct | openai/gsm8k | [](https://colab.research.google.com/drive/11tj2odJa7v55VvQkpOlm7_WBclqFdaVi?usp=sharing) |
|
|
202
|
+
| TRL | **GRPO** | meta-llama/Llama-3.2-3B-Instruct | openai/gsm8k | [](https://colab.research.google.com/drive/14b2dt0_iSVL8Z_8mlOx--f-d_rDCqVOP?usp=sharing) |
|
|
203
|
+
| Unsloth | **DRGRPO** | Qwen/Qwen2.5-3B-Instruct | yahma/alpaca-cleaned | [](https://colab.research.google.com/drive/1IzDtjONyL6CZ038faTTp8AvLfhzkqSpb?usp=sharing) |
|
|
204
|
+
| TRL | **DRGRPO** | Qwen/Qwen2-0.5B-Instruct | AI-MO/NuminaMath-TIR | [](https://colab.research.google.com/drive/1KfrMzuGRPrZwyRuNTGep1ZL6C6h2aVjk?usp=sharing) |
|
|
205
|
+
| Unsloth | **GSPO** | Qwen/Qwen3-1.7B | CyberNative/Code_Vulnerability_Security_DPO | [](https://colab.research.google.com/drive/120FjMEAZRXUoOMRsWHCn5tmpGfLTAY4g?usp=sharing) |
|
|
206
|
+
| TRL | **GSPO** | meta-llama/Llama-3.2-3B-Instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1vDP7ukBHWwSiD7KVSgSekCvk-x_MkWdC?usp=sharing) |
|
|
207
|
+
| Unsloth| **DAPO** | microsoft/Phi-3.5-mini-instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1gAa6EPc5345XRfog1nzQIFTmp_OJqNvk?usp=sharing) |
|
|
208
|
+
|TRL | **DAPO** | meta-llama/Llama-3.2-3B-Instruct | google-research-datasets/mbpp | [](https://colab.research.google.com/drive/1aF3LxEMmLl0fCyA5yVBAkU_Rgsy3dsh5?usp=sharing) |
|
|
209
|
+
## Documentation
|
|
210
|
+
|
|
211
|
+
- **[Getting Started](docs/getting-started/installation.md)**: Installation, setup, and basic usage
|
|
212
|
+
- **[User Guide](docs/user-guide/overview.md)**: In-depth tutorials for SFT and RL training
|
|
213
|
+
- **[API Reference](docs/api-reference/overview.md)**: Complete Python API and class/method details
|
|
214
|
+
- **[Examples](docs/examples/overview.md)**: End-to-end code examples
|
|
215
|
+
- **[Advanced Topics](docs/advanced/architecture.md)**: Architecture, custom backends, and performance optimization
|
|
216
|
+
- **[Notebooks](docs/notebooks/index.md)**: Interactive Colab notebooks and local Jupyter notebooks
|
|
217
|
+
|
|
218
|
+
## Key Capabilities
|
|
219
|
+
|
|
220
|
+
- **Multiple Training Paradigms**: Supports SFT, DPO, PPO, GRPO, and advanced RL algorithms
|
|
221
|
+
- **Backend Flexibility**: TRL and Unsloth backends with automatic fallback
|
|
222
|
+
- **Reward Model Training**: Train custom reward models from rule-based functions
|
|
223
|
+
- **Comprehensive Evaluation**: Multi-level evaluation with lm-eval integration
|
|
224
|
+
- **Production Ready**: Model serialization, reproducible training, and deployment-ready pipelines
|
|
225
|
+
- **Extensible Architecture**: Modular design for easy integration of custom algorithms and backends
|
|
226
|
+
|
|
227
|
+
## Architecture
|
|
228
|
+
|
|
229
|
+
AlignTune uses a flexible backend architecture:
|
|
230
|
+
|
|
231
|
+
```mermaid
|
|
232
|
+
flowchart TD
|
|
233
|
+
Factory[Backend Factory] --> TRL[TRL Backend]
|
|
234
|
+
Factory --> Unsloth[Unsloth Backend]
|
|
235
|
+
TRL --> TRL_Algos[TRL Algorithms]
|
|
236
|
+
Unsloth --> Unsloth_Algos[Unsloth Algorithms]
|
|
237
|
+
```
|
|
238
|
+
|
|
239
|
+
**TRL Backend:** SFT, DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, PACE
|
|
240
|
+
|
|
241
|
+
**Unsloth Backend:** SFT, DPO, PPO, GRPO, DAPO, Dr. GRPO, Counterfactual GRPO, PACE
|
|
242
|
+
|
|
243
|
+
See [Architecture](docs/advanced/architecture.md) for details.
|
|
244
|
+
|
|
245
|
+
## Contributing
|
|
246
|
+
|
|
247
|
+
We welcome contributions! See our [Contributing Guide](docs/contributing/guide.md) for details.
|
|
248
|
+
|
|
249
|
+
## License
|
|
250
|
+
|
|
251
|
+
This project is released under the MIT License. Please cite appropriately if used in academic or production projects.
|
|
252
|
+
See the [LICENSE](LICENSE) file for details.
|
|
253
|
+
|
|
254
|
+
**Key Points:**
|
|
255
|
+
- **Free for Research & Learning**: Use, modify, and study for personal, academic, or research purposes
|
|
256
|
+
- **Source Available**: Full access to source code
|
|
257
|
+
- **Commercial Use Restricted**: Requires separate commercial license
|
|
258
|
+
- **Contact**: For commercial licensing, partnership, or redistribution rights, contact [support@lexsi.ai](mailto:support@lexsi.ai)
|
|
259
|
+
|
|
260
|
+
This is **not** an open-source license as defined by OSI, but provides broad access for non-commercial use.
|
|
261
|
+
|
|
262
|
+
## Citation
|
|
263
|
+
|
|
264
|
+
If you use AlignTune in your research, please cite:
|
|
265
|
+
|
|
266
|
+
**BibTeX:**
|
|
267
|
+
```bibtex
|
|
268
|
+
@software{alignTune2025,
|
|
269
|
+
title = {{AlignTune}: Modular Toolkit for Post-Training Alignment of Large Language Models},
|
|
270
|
+
author = {Lyngkhoi, R E Zera Marveen and Chawla, Chirag and Seth, Pratinav and Avaiya, Utsav and Bhattacharjee, Soham and Khandoga, Mykola and Yuan, Rui and Sankarapu, Vinay Kumar},
|
|
271
|
+
year = {2025},
|
|
272
|
+
note = {Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth},
|
|
273
|
+
organization = {Lexsi Labs},
|
|
274
|
+
url = {https://github.com/Lexsi-Labs/aligntune},
|
|
275
|
+
version = {0.0.0}
|
|
276
|
+
}
|
|
277
|
+
```
|
|
278
|
+
|
|
279
|
+
**Plain Text:**
|
|
280
|
+
```
|
|
281
|
+
Lyngkhoi, R. E. Z. M., Chawla, C., Seth, P., Avaiya, U., Bhattacharjee, S., Khandoga, M.,
|
|
282
|
+
Yuan, R., & Sankarapu, V. K. (2025). AlignTune: Modular Toolkit for Post-Training Alignment
|
|
283
|
+
of Large Language Models. Lexsi Labs. https://github.com/Lexsi-Labs/aligntune
|
|
284
|
+
|
|
285
|
+
*Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth
|
|
286
|
+
```
|
|
287
|
+
|
|
288
|
+
## Acknowledgments
|
|
289
|
+
|
|
290
|
+
AlignTune is built upon the excellent work of the following projects:
|
|
291
|
+
|
|
292
|
+
- **[HuggingFace Transformers](https://github.com/huggingface/transformers)** - Model architectures and tokenizers
|
|
293
|
+
- **[TRL](https://github.com/huggingface/trl)** - Transformer Reinforcement Learning library
|
|
294
|
+
- **[Unsloth](https://github.com/unslothai/unsloth)** - Fast and memory-efficient training
|
|
295
|
+
- **[HuggingFace Datasets](https://github.com/huggingface/datasets)** - Dataset loading and processing
|
|
296
|
+
|
|
297
|
+
## Support
|
|
298
|
+
|
|
299
|
+
- **Documentation**: [aligntune.lexsi.ai/](https://aligntune.lexsi.ai/)
|
|
300
|
+
- **GitHub Issues**: [github.com/Lexsi-Labs/aligntune/issues](https://github.com/Lexsi-Labs/aligntune/issues)
|
|
301
|
+
- **Discussions**: [github.com/Lexsi-Labs/aligntune/discussions](https://github.com/Lexsi-Labs/aligntune/discussions)
|
|
302
|
+
- **Email**: [support@lexsi.ai](mailto:support@lexsi.ai)
|
|
303
|
+
|
|
304
|
+
---
|
|
305
|
+
|
|
306
|
+
**Get started with AlignTune and accelerate your LLM fine-tuning workflows today!**
|
|
@@ -0,0 +1,230 @@
|
|
|
1
|
+
|
|
2
|
+
<p align="center">
|
|
3
|
+
<img src="docs/assets/aligntune-banner.png" alt="AlignTune Banner" width="1000px"/>
|
|
4
|
+
</p>
|
|
5
|
+
|
|
6
|
+
<div align="center">
|
|
7
|
+
<a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.8+-blue.svg"/></a>
|
|
8
|
+
<a href="https://github.com/Lexsi-Labs/aligntune/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg"/></a>
|
|
9
|
+
<a href="https://badge.fury.io/py/aligntune"><img src="https://badge.fury.io/py/aligntune.svg"/></a>
|
|
10
|
+
</div>
|
|
11
|
+
|
|
12
|
+
---
|
|
13
|
+
|
|
14
|
+
**AlignTune** is a production-ready fine-tuning library designed to simplify training and fine-tuning of Large Language Models (LLMs) with both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) methods. It provides a high-level, unified API that abstracts away the complexities of backend selection, algorithm configuration, and training loops, letting you focus on delivering results.
|
|
15
|
+
|
|
16
|
+
## Core Features
|
|
17
|
+
|
|
18
|
+
**Multi-Backend Architecture**: Choose between TRL (reliable, battle-tested) and Unsloth (faster) backends with intelligent auto-selection.
|
|
19
|
+
|
|
20
|
+
**Complete RLHF Coverage**: 12+ RL algorithms including DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, and PACE.
|
|
21
|
+
|
|
22
|
+
**Production-Ready**: No mock code, comprehensive error handling, extensive testing, and robust validation.
|
|
23
|
+
|
|
24
|
+
## Quick Start
|
|
25
|
+
|
|
26
|
+
### Supervised Fine-Tuning (SFT)
|
|
27
|
+
|
|
28
|
+
```python
|
|
29
|
+
from aligntune.core.backend_factory import create_sft_trainer
|
|
30
|
+
|
|
31
|
+
# Create and train SFT model
|
|
32
|
+
trainer = create_sft_trainer(
|
|
33
|
+
model_name="microsoft/DialoGPT-small",
|
|
34
|
+
dataset_name="tatsu-lab/alpaca",
|
|
35
|
+
backend="trl",
|
|
36
|
+
num_epochs=3,
|
|
37
|
+
max_steps = -1,
|
|
38
|
+
batch_size=4,
|
|
39
|
+
learning_rate=5e-5
|
|
40
|
+
)
|
|
41
|
+
|
|
42
|
+
# Train the model
|
|
43
|
+
trainer.train()
|
|
44
|
+
|
|
45
|
+
# Evaluate
|
|
46
|
+
metrics = trainer.evaluate()
|
|
47
|
+
print(metrics)
|
|
48
|
+
```
|
|
49
|
+
|
|
50
|
+
### Reinforcement Learning (DPO)
|
|
51
|
+
|
|
52
|
+
```python
|
|
53
|
+
from aligntune.core.backend_factory import create_rl_trainer
|
|
54
|
+
|
|
55
|
+
# Create and train DPO model
|
|
56
|
+
trainer = create_rl_trainer(
|
|
57
|
+
model_name="Qwen/Qwen3-0.6B",
|
|
58
|
+
dataset_name="Anthropic/hh-rlhf",
|
|
59
|
+
algorithm="dpo",
|
|
60
|
+
backend="trl",
|
|
61
|
+
num_epochs=1,
|
|
62
|
+
batch_size=4,
|
|
63
|
+
learning_rate=5e-5
|
|
64
|
+
)
|
|
65
|
+
|
|
66
|
+
# Train the model
|
|
67
|
+
trainer.train()
|
|
68
|
+
```
|
|
69
|
+
|
|
70
|
+
## Supported Algorithms
|
|
71
|
+
|
|
72
|
+
| Algorithm | TRL Backend | Unsloth Backend | Description |
|
|
73
|
+
|-----------|-------------|-----------------|-------------|
|
|
74
|
+
| **SFT** | Yes | Yes | Supervised Fine-Tuning |
|
|
75
|
+
| **DPO** | Yes | Yes | Direct Preference Optimization |
|
|
76
|
+
| **PPO** | Yes | Yes | Proximal Policy Optimization |
|
|
77
|
+
| **GRPO** | Yes | Yes | Group Relative Policy Optimization |
|
|
78
|
+
| **GSPO** | Yes | Yes | Group Sequential Policy Optimization |
|
|
79
|
+
| **DAPO** | Yes | Yes | Decouple Clip and Dynamic sAmpling Policy Optimization |
|
|
80
|
+
| **Dr. GRPO** | Yes | Yes | GRPO Done Right (unbiased variant) |
|
|
81
|
+
| **GBMPO** | Yes | No | Group-Based Mirror Policy Optimization |
|
|
82
|
+
| **Counterfactual GRPO** | Yes | Yes | Counterfactual GRPO variant |
|
|
83
|
+
| **PACE** | Yes | Yes | Baseline-Optimized Learning Technique |
|
|
84
|
+
|
|
85
|
+
## Installation
|
|
86
|
+
|
|
87
|
+
```bash
|
|
88
|
+
# Or install from source
|
|
89
|
+
git clone https://github.com/Lexsi-Labs/aligntune.git
|
|
90
|
+
cd aligntune
|
|
91
|
+
pip install -e .
|
|
92
|
+
```
|
|
93
|
+
|
|
94
|
+
### Requirements
|
|
95
|
+
|
|
96
|
+
- Python 3.8+
|
|
97
|
+
- PyTorch 2.0+
|
|
98
|
+
- CUDA-compatible GPU (recommended for faster training)
|
|
99
|
+
|
|
100
|
+
## Demo Notebooks
|
|
101
|
+
|
|
102
|
+
Interactive Colab notebooks demonstrating various AlignTune workflows:
|
|
103
|
+
Here are the organized tables containing the Colab links, models, and datasets provided in your text.
|
|
104
|
+
|
|
105
|
+
### Supervised Fine-Tuning (SFT)
|
|
106
|
+
|
|
107
|
+
| Backend| Model | Dataset | Link |
|
|
108
|
+
| --- | --- | --- | --- |
|
|
109
|
+
| TRL | **Qwen/Qwen3-4B-Instruct-2507** | sohamb37lexsi/bitext-wealth-management-llm-chatbot-splits | [](https://colab.research.google.com/drive/1okAnfMlkch-G5Dy2dmj_dfR07rtxmYoq?usp=sharing) |
|
|
110
|
+
| TRL | **Qwen3-4B-Instruct** | sohamb37lexsi/bitext-retail-banking-llm-chatbot-splits | [](https://colab.research.google.com/drive/1JnFvEWZ6PDqrDAyWznCdOA9oNn8aqN2v?usp=sharing) |
|
|
111
|
+
| Unsloth| **Qwen/Qwen2.5-0.5B-Instruct** | bebechien/MobileGameNPC | [](https://colab.research.google.com/drive/15R3JOrzAUuMagCamDHsqe0wNyb44rw2j?usp=sharing) |
|
|
112
|
+
| TRL | **google/txgemma-2b-predict** | trialbench_adverse-event-rate-prediction | [](https://colab.research.google.com/drive/1r91SS7Lb5LjkjzWXp7uzkpR3vlWZ6gwD?usp=sharing) |
|
|
113
|
+
| Unsloth| **Qwen/Qwen2.5-0.5B-Instruct** | bebechien/MobileGameNP |[](https://colab.research.google.com/drive/11DrRbG30MUCegZTDXwR9nxEdvKaLfWCb?usp=sharing) |
|
|
114
|
+
|
|
115
|
+
### Reinforcement Learning (RL)
|
|
116
|
+
|
|
117
|
+
| Backend| Algorithm | Model | Dataset | Link |
|
|
118
|
+
| --- | --- | --- | --- | --- |
|
|
119
|
+
| Unsloth| **DPO** | microsoft/phi-2 | argilla/distilabel-intel-orca-dpo-pairs | [](https://colab.research.google.com/drive/1aKdQzT7KGs0PEr7pO9rOXFkwQK6LzyRZ#scrollTo=dOnIJIAMxP9J) |
|
|
120
|
+
| TRL | **DPO** | google/gemma-2-2b-it | Anthropic/hh-rlhf | [](https://colab.research.google.com/drive/1WQhek61Z0v1zHFWT4mQwVX7n10ln-j3d?usp=sharing) |
|
|
121
|
+
| TRL | **DPO** | sohamb37lexsi/wealth_management_Qwen3-4B-Instruct-2507 | sohamb37lexsi/bitext_wealth_management_preference_data | [](https://colab.research.google.com/drive/1LR3KWcjQAFjSHf0MR9271ShchjPUVSpS?usp=sharing) |
|
|
122
|
+
| Unsloth| **PPO** | Qwen/Qwen2.5-0.5B-Instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1G4fdtO4DhBNwjOYhuaGmvIrBU_lc159H?usp=sharing) |
|
|
123
|
+
| TRL | **PPO** | EleutherAI/pythia-1.4b | CarperAI/openai_summarize_tldr | [](https://colab.research.google.com/drive/1wlbSvQAJv8ZFM2qGD4XlUOKPzZHXstQo?usp=sharing) |
|
|
124
|
+
| TRL | **GRPO** (Coding) | Qwen/Qwen3-4B | google-research-datasets/mbpp | [](https://colab.research.google.com/drive/13HYZ-EkLC3-6wxJG_1NyWXPEuN9bIeM4?usp=sharing) |
|
|
125
|
+
| Unsloth| **GRPO** (Math) | meta-llama/Llama-3.2-3B-Instruct | openai/gsm8k | [](https://colab.research.google.com/drive/11tj2odJa7v55VvQkpOlm7_WBclqFdaVi?usp=sharing) |
|
|
126
|
+
| TRL | **GRPO** | meta-llama/Llama-3.2-3B-Instruct | openai/gsm8k | [](https://colab.research.google.com/drive/14b2dt0_iSVL8Z_8mlOx--f-d_rDCqVOP?usp=sharing) |
|
|
127
|
+
| Unsloth | **DRGRPO** | Qwen/Qwen2.5-3B-Instruct | yahma/alpaca-cleaned | [](https://colab.research.google.com/drive/1IzDtjONyL6CZ038faTTp8AvLfhzkqSpb?usp=sharing) |
|
|
128
|
+
| TRL | **DRGRPO** | Qwen/Qwen2-0.5B-Instruct | AI-MO/NuminaMath-TIR | [](https://colab.research.google.com/drive/1KfrMzuGRPrZwyRuNTGep1ZL6C6h2aVjk?usp=sharing) |
|
|
129
|
+
| Unsloth | **GSPO** | Qwen/Qwen3-1.7B | CyberNative/Code_Vulnerability_Security_DPO | [](https://colab.research.google.com/drive/120FjMEAZRXUoOMRsWHCn5tmpGfLTAY4g?usp=sharing) |
|
|
130
|
+
| TRL | **GSPO** | meta-llama/Llama-3.2-3B-Instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1vDP7ukBHWwSiD7KVSgSekCvk-x_MkWdC?usp=sharing) |
|
|
131
|
+
| Unsloth| **DAPO** | microsoft/Phi-3.5-mini-instruct | HuggingFaceH4/ultrachat_200k | [](https://colab.research.google.com/drive/1gAa6EPc5345XRfog1nzQIFTmp_OJqNvk?usp=sharing) |
|
|
132
|
+
|TRL | **DAPO** | meta-llama/Llama-3.2-3B-Instruct | google-research-datasets/mbpp | [](https://colab.research.google.com/drive/1aF3LxEMmLl0fCyA5yVBAkU_Rgsy3dsh5?usp=sharing) |
|
|
133
|
+
## Documentation
|
|
134
|
+
|
|
135
|
+
- **[Getting Started](docs/getting-started/installation.md)**: Installation, setup, and basic usage
|
|
136
|
+
- **[User Guide](docs/user-guide/overview.md)**: In-depth tutorials for SFT and RL training
|
|
137
|
+
- **[API Reference](docs/api-reference/overview.md)**: Complete Python API and class/method details
|
|
138
|
+
- **[Examples](docs/examples/overview.md)**: End-to-end code examples
|
|
139
|
+
- **[Advanced Topics](docs/advanced/architecture.md)**: Architecture, custom backends, and performance optimization
|
|
140
|
+
- **[Notebooks](docs/notebooks/index.md)**: Interactive Colab notebooks and local Jupyter notebooks
|
|
141
|
+
|
|
142
|
+
## Key Capabilities
|
|
143
|
+
|
|
144
|
+
- **Multiple Training Paradigms**: Supports SFT, DPO, PPO, GRPO, and advanced RL algorithms
|
|
145
|
+
- **Backend Flexibility**: TRL and Unsloth backends with automatic fallback
|
|
146
|
+
- **Reward Model Training**: Train custom reward models from rule-based functions
|
|
147
|
+
- **Comprehensive Evaluation**: Multi-level evaluation with lm-eval integration
|
|
148
|
+
- **Production Ready**: Model serialization, reproducible training, and deployment-ready pipelines
|
|
149
|
+
- **Extensible Architecture**: Modular design for easy integration of custom algorithms and backends
|
|
150
|
+
|
|
151
|
+
## Architecture
|
|
152
|
+
|
|
153
|
+
AlignTune uses a flexible backend architecture:
|
|
154
|
+
|
|
155
|
+
```mermaid
|
|
156
|
+
flowchart TD
|
|
157
|
+
Factory[Backend Factory] --> TRL[TRL Backend]
|
|
158
|
+
Factory --> Unsloth[Unsloth Backend]
|
|
159
|
+
TRL --> TRL_Algos[TRL Algorithms]
|
|
160
|
+
Unsloth --> Unsloth_Algos[Unsloth Algorithms]
|
|
161
|
+
```
|
|
162
|
+
|
|
163
|
+
**TRL Backend:** SFT, DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, PACE
|
|
164
|
+
|
|
165
|
+
**Unsloth Backend:** SFT, DPO, PPO, GRPO, DAPO, Dr. GRPO, Counterfactual GRPO, PACE
|
|
166
|
+
|
|
167
|
+
See [Architecture](docs/advanced/architecture.md) for details.
|
|
168
|
+
|
|
169
|
+
## Contributing
|
|
170
|
+
|
|
171
|
+
We welcome contributions! See our [Contributing Guide](docs/contributing/guide.md) for details.
|
|
172
|
+
|
|
173
|
+
## License
|
|
174
|
+
|
|
175
|
+
This project is released under the MIT License. Please cite appropriately if used in academic or production projects.
|
|
176
|
+
See the [LICENSE](LICENSE) file for details.
|
|
177
|
+
|
|
178
|
+
**Key Points:**
|
|
179
|
+
- **Free for Research & Learning**: Use, modify, and study for personal, academic, or research purposes
|
|
180
|
+
- **Source Available**: Full access to source code
|
|
181
|
+
- **Commercial Use Restricted**: Requires separate commercial license
|
|
182
|
+
- **Contact**: For commercial licensing, partnership, or redistribution rights, contact [support@lexsi.ai](mailto:support@lexsi.ai)
|
|
183
|
+
|
|
184
|
+
This is **not** an open-source license as defined by OSI, but provides broad access for non-commercial use.
|
|
185
|
+
|
|
186
|
+
## Citation
|
|
187
|
+
|
|
188
|
+
If you use AlignTune in your research, please cite:
|
|
189
|
+
|
|
190
|
+
**BibTeX:**
|
|
191
|
+
```bibtex
|
|
192
|
+
@software{alignTune2025,
|
|
193
|
+
title = {{AlignTune}: Modular Toolkit for Post-Training Alignment of Large Language Models},
|
|
194
|
+
author = {Lyngkhoi, R E Zera Marveen and Chawla, Chirag and Seth, Pratinav and Avaiya, Utsav and Bhattacharjee, Soham and Khandoga, Mykola and Yuan, Rui and Sankarapu, Vinay Kumar},
|
|
195
|
+
year = {2025},
|
|
196
|
+
note = {Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth},
|
|
197
|
+
organization = {Lexsi Labs},
|
|
198
|
+
url = {https://github.com/Lexsi-Labs/aligntune},
|
|
199
|
+
version = {0.0.0}
|
|
200
|
+
}
|
|
201
|
+
```
|
|
202
|
+
|
|
203
|
+
**Plain Text:**
|
|
204
|
+
```
|
|
205
|
+
Lyngkhoi, R. E. Z. M., Chawla, C., Seth, P., Avaiya, U., Bhattacharjee, S., Khandoga, M.,
|
|
206
|
+
Yuan, R., & Sankarapu, V. K. (2025). AlignTune: Modular Toolkit for Post-Training Alignment
|
|
207
|
+
of Large Language Models. Lexsi Labs. https://github.com/Lexsi-Labs/aligntune
|
|
208
|
+
|
|
209
|
+
*Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth
|
|
210
|
+
```
|
|
211
|
+
|
|
212
|
+
## Acknowledgments
|
|
213
|
+
|
|
214
|
+
AlignTune is built upon the excellent work of the following projects:
|
|
215
|
+
|
|
216
|
+
- **[HuggingFace Transformers](https://github.com/huggingface/transformers)** - Model architectures and tokenizers
|
|
217
|
+
- **[TRL](https://github.com/huggingface/trl)** - Transformer Reinforcement Learning library
|
|
218
|
+
- **[Unsloth](https://github.com/unslothai/unsloth)** - Fast and memory-efficient training
|
|
219
|
+
- **[HuggingFace Datasets](https://github.com/huggingface/datasets)** - Dataset loading and processing
|
|
220
|
+
|
|
221
|
+
## Support
|
|
222
|
+
|
|
223
|
+
- **Documentation**: [aligntune.lexsi.ai/](https://aligntune.lexsi.ai/)
|
|
224
|
+
- **GitHub Issues**: [github.com/Lexsi-Labs/aligntune/issues](https://github.com/Lexsi-Labs/aligntune/issues)
|
|
225
|
+
- **Discussions**: [github.com/Lexsi-Labs/aligntune/discussions](https://github.com/Lexsi-Labs/aligntune/discussions)
|
|
226
|
+
- **Email**: [support@lexsi.ai](mailto:support@lexsi.ai)
|
|
227
|
+
|
|
228
|
+
---
|
|
229
|
+
|
|
230
|
+
**Get started with AlignTune and accelerate your LLM fine-tuning workflows today!**
|