aria-trl 1.0.0__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.
- aria_trl-1.0.0/.gitignore +146 -0
- aria_trl-1.0.0/LICENSE +21 -0
- aria_trl-1.0.0/PKG-INFO +301 -0
- aria_trl-1.0.0/README.md +259 -0
- aria_trl-1.0.0/aria_trl/__init__.py +22 -0
- aria_trl-1.0.0/aria_trl/config.py +41 -0
- aria_trl-1.0.0/aria_trl/consolidation.py +147 -0
- aria_trl-1.0.0/aria_trl/modules.py +167 -0
- aria_trl-1.0.0/aria_trl/trainer.py +264 -0
- aria_trl-1.0.0/aria_trl/utils.py +130 -0
- aria_trl-1.0.0/aria_trl.egg-info/PKG-INFO +301 -0
- aria_trl-1.0.0/aria_trl.egg-info/SOURCES.txt +18 -0
- aria_trl-1.0.0/aria_trl.egg-info/dependency_links.txt +1 -0
- aria_trl-1.0.0/aria_trl.egg-info/requires.txt +14 -0
- aria_trl-1.0.0/aria_trl.egg-info/top_level.txt +1 -0
- aria_trl-1.0.0/examples/__init__.py +1 -0
- aria_trl-1.0.0/examples/distilgpt2_example.py +237 -0
- aria_trl-1.0.0/pyproject.toml +74 -0
- aria_trl-1.0.0/setup.cfg +4 -0
- aria_trl-1.0.0/setup.py +47 -0
|
@@ -0,0 +1,146 @@
|
|
|
1
|
+
# Byte-compiled / optimized / DLL files
|
|
2
|
+
__pycache__/
|
|
3
|
+
*.py[cod]
|
|
4
|
+
*$py.class
|
|
5
|
+
|
|
6
|
+
# C extensions
|
|
7
|
+
*.so
|
|
8
|
+
|
|
9
|
+
# Distribution / packaging
|
|
10
|
+
.Python
|
|
11
|
+
build/
|
|
12
|
+
develop-eggs/
|
|
13
|
+
dist/
|
|
14
|
+
downloads/
|
|
15
|
+
eggs/
|
|
16
|
+
.eggs/
|
|
17
|
+
lib/
|
|
18
|
+
lib64/
|
|
19
|
+
parts/
|
|
20
|
+
sdist/
|
|
21
|
+
var/
|
|
22
|
+
wheels/
|
|
23
|
+
pip-wheel-metadata/
|
|
24
|
+
share/python-wheels/
|
|
25
|
+
*.egg-info/
|
|
26
|
+
.installed.cfg
|
|
27
|
+
*.egg
|
|
28
|
+
MANIFEST
|
|
29
|
+
|
|
30
|
+
# PyInstaller
|
|
31
|
+
*.manifest
|
|
32
|
+
*.spec
|
|
33
|
+
|
|
34
|
+
# Installer logs
|
|
35
|
+
pip-log.txt
|
|
36
|
+
pip-delete-this-directory.txt
|
|
37
|
+
|
|
38
|
+
# Unit test / coverage reports
|
|
39
|
+
htmlcov/
|
|
40
|
+
.tox/
|
|
41
|
+
.nox/
|
|
42
|
+
.coverage
|
|
43
|
+
.coverage.*
|
|
44
|
+
.cache
|
|
45
|
+
nosetests.xml
|
|
46
|
+
coverage.xml
|
|
47
|
+
*.cover
|
|
48
|
+
*.py,cover
|
|
49
|
+
.hypothesis/
|
|
50
|
+
.pytest_cache/
|
|
51
|
+
|
|
52
|
+
# Translations
|
|
53
|
+
*.mo
|
|
54
|
+
*.pot
|
|
55
|
+
|
|
56
|
+
# Django stuff:
|
|
57
|
+
*.log
|
|
58
|
+
local_settings.py
|
|
59
|
+
db.sqlite3
|
|
60
|
+
db.sqlite3-journal
|
|
61
|
+
|
|
62
|
+
# Flask stuff:
|
|
63
|
+
instance/
|
|
64
|
+
.webassets-cache
|
|
65
|
+
|
|
66
|
+
# Scrapy stuff:
|
|
67
|
+
.scrapy
|
|
68
|
+
|
|
69
|
+
# Sphinx documentation
|
|
70
|
+
docs/_build/
|
|
71
|
+
|
|
72
|
+
# PyBuilder
|
|
73
|
+
target/
|
|
74
|
+
|
|
75
|
+
# Jupyter Notebook
|
|
76
|
+
.ipynb_checkpoints
|
|
77
|
+
|
|
78
|
+
# IPython
|
|
79
|
+
profile_default/
|
|
80
|
+
ipython_config.py
|
|
81
|
+
|
|
82
|
+
# pyenv
|
|
83
|
+
.python-version
|
|
84
|
+
|
|
85
|
+
# pipenv
|
|
86
|
+
Pipfile.lock
|
|
87
|
+
|
|
88
|
+
# PEP 582
|
|
89
|
+
__pypackages__/
|
|
90
|
+
|
|
91
|
+
# Celery stuff
|
|
92
|
+
celerybeat-schedule
|
|
93
|
+
celerybeat.pid
|
|
94
|
+
|
|
95
|
+
# SageMath parsed files
|
|
96
|
+
*.sage.py
|
|
97
|
+
|
|
98
|
+
# Environments
|
|
99
|
+
.env
|
|
100
|
+
.venv
|
|
101
|
+
env/
|
|
102
|
+
venv/
|
|
103
|
+
ENV/
|
|
104
|
+
env.bak/
|
|
105
|
+
venv.bak/
|
|
106
|
+
|
|
107
|
+
# Spyder project settings
|
|
108
|
+
.spyderproject
|
|
109
|
+
.spyproject
|
|
110
|
+
|
|
111
|
+
# Rope project settings
|
|
112
|
+
.ropeproject
|
|
113
|
+
|
|
114
|
+
# mkdocs documentation
|
|
115
|
+
/site
|
|
116
|
+
|
|
117
|
+
# mypy
|
|
118
|
+
.mypy_cache/
|
|
119
|
+
.dmypy.json
|
|
120
|
+
dmypy.json
|
|
121
|
+
|
|
122
|
+
# Pyre type checker
|
|
123
|
+
.pyre/
|
|
124
|
+
|
|
125
|
+
# IDE
|
|
126
|
+
.vscode/
|
|
127
|
+
.idea/
|
|
128
|
+
*.swp
|
|
129
|
+
*.swo
|
|
130
|
+
*~
|
|
131
|
+
.DS_Store
|
|
132
|
+
|
|
133
|
+
# Model checkpoints and results
|
|
134
|
+
checkpoints/
|
|
135
|
+
results/
|
|
136
|
+
*.pt
|
|
137
|
+
*.pth
|
|
138
|
+
*.bin
|
|
139
|
+
|
|
140
|
+
# Tensorboard
|
|
141
|
+
runs/
|
|
142
|
+
logs/
|
|
143
|
+
|
|
144
|
+
# Temporary files
|
|
145
|
+
*.tmp
|
|
146
|
+
*.temp
|
aria_trl-1.0.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Darshan Poudel
|
|
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.
|
aria_trl-1.0.0/PKG-INFO
ADDED
|
@@ -0,0 +1,301 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: aria-trl
|
|
3
|
+
Version: 1.0.0
|
|
4
|
+
Summary: Continual learning for LLM fine-tuning via ARIA mechanisms (PlasticityGatedMLP, SPC, Task Adapters)
|
|
5
|
+
Home-page: https://github.com/rsd-darshan/aria-trl
|
|
6
|
+
Author: Darshan Poudel
|
|
7
|
+
Author-email: Darshan Poudel <poudeldarshan44@gmail.com>
|
|
8
|
+
License: MIT
|
|
9
|
+
Project-URL: Homepage, https://github.com/rsd-darshan/aria-trl
|
|
10
|
+
Project-URL: Repository, https://github.com/rsd-darshan/aria-trl.git
|
|
11
|
+
Project-URL: Issues, https://github.com/rsd-darshan/aria-trl/issues
|
|
12
|
+
Keywords: continual-learning,llm,fine-tuning,catastrophic-forgetting,fast-slow-pathways,fisher-consolidation
|
|
13
|
+
Classifier: Development Status :: 4 - Beta
|
|
14
|
+
Classifier: Intended Audience :: Developers
|
|
15
|
+
Classifier: Intended Audience :: Science/Research
|
|
16
|
+
Classifier: License :: OSI Approved :: MIT License
|
|
17
|
+
Classifier: Operating System :: OS Independent
|
|
18
|
+
Classifier: Programming Language :: Python :: 3
|
|
19
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
20
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
21
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
22
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
23
|
+
Requires-Python: >=3.10
|
|
24
|
+
Description-Content-Type: text/markdown
|
|
25
|
+
License-File: LICENSE
|
|
26
|
+
Requires-Dist: torch>=2.0.0
|
|
27
|
+
Requires-Dist: transformers>=4.30.0
|
|
28
|
+
Requires-Dist: trl>=0.7.0
|
|
29
|
+
Requires-Dist: datasets>=2.10.0
|
|
30
|
+
Requires-Dist: numpy>=1.23.0
|
|
31
|
+
Provides-Extra: dev
|
|
32
|
+
Requires-Dist: pytest>=7.0; extra == "dev"
|
|
33
|
+
Requires-Dist: black>=23.0; extra == "dev"
|
|
34
|
+
Requires-Dist: isort>=5.12; extra == "dev"
|
|
35
|
+
Requires-Dist: mypy>=1.0; extra == "dev"
|
|
36
|
+
Provides-Extra: examples
|
|
37
|
+
Requires-Dist: wandb>=0.15.0; extra == "examples"
|
|
38
|
+
Dynamic: author
|
|
39
|
+
Dynamic: home-page
|
|
40
|
+
Dynamic: license-file
|
|
41
|
+
Dynamic: requires-python
|
|
42
|
+
|
|
43
|
+
# aria-trl: Continual Learning for LLM Fine-tuning
|
|
44
|
+
|
|
45
|
+
[](https://www.python.org/downloads/)
|
|
46
|
+
[](https://pytorch.org/)
|
|
47
|
+
[](LICENSE)
|
|
48
|
+
|
|
49
|
+
**aria-trl** brings ARIA's continual learning mechanisms to Hugging Face's TRL library, enabling fine-tuning of large language models on sequential tasks without catastrophic forgetting.
|
|
50
|
+
|
|
51
|
+
## Overview
|
|
52
|
+
|
|
53
|
+
Fine-tuning LLMs sequentially on multiple tasks typically leads to **catastrophic forgetting** — the model forgets earlier tasks while learning new ones. aria-trl prevents this through three ARIA mechanisms:
|
|
54
|
+
|
|
55
|
+
1. **PlasticityGatedMLP**: Dual fast/slow pathways in FFN layers
|
|
56
|
+
- Fast pathway: volatile, learns task-specific patterns
|
|
57
|
+
- Slow pathway: stable, retains task-generic knowledge
|
|
58
|
+
- Learned gate routes computation per token
|
|
59
|
+
|
|
60
|
+
2. **Slow-Pathway Consolidation (SPC)**: Fisher-weighted regularization
|
|
61
|
+
- Estimates diagonal Fisher Information after each task
|
|
62
|
+
- Protects consolidated knowledge from being overwritten
|
|
63
|
+
- 50% fewer parameters than standard EWC
|
|
64
|
+
|
|
65
|
+
3. **Task-Specific Adapters**: Lightweight LoRA-like modules
|
|
66
|
+
- One residual adapter per task
|
|
67
|
+
- Frozen after training to prevent overwriting
|
|
68
|
+
- Minimal parameter overhead
|
|
69
|
+
|
|
70
|
+
## Features
|
|
71
|
+
|
|
72
|
+
- ✅ **Drop-in SFTTrainer subclass** — minimal code changes
|
|
73
|
+
- ✅ **Asymmetric learning rates** — slow pathway learns slower
|
|
74
|
+
- ✅ **Gradient dampening** — slow grads scaled by (1−π̄)
|
|
75
|
+
- ✅ **Works with any HF model** — LLaMA, Mistral, DistilGPT2, etc.
|
|
76
|
+
- ✅ **Full checkpoint support** — save/load with consolidation state
|
|
77
|
+
- ✅ **Production-ready** — comprehensive error handling
|
|
78
|
+
|
|
79
|
+
## Installation
|
|
80
|
+
|
|
81
|
+
```bash
|
|
82
|
+
pip install aria-trl
|
|
83
|
+
```
|
|
84
|
+
|
|
85
|
+
Or install from source:
|
|
86
|
+
|
|
87
|
+
```bash
|
|
88
|
+
git clone https://github.com/rsd-darshan/aria-trl.git
|
|
89
|
+
cd aria-trl
|
|
90
|
+
pip install -e .
|
|
91
|
+
```
|
|
92
|
+
|
|
93
|
+
## Quick Start
|
|
94
|
+
|
|
95
|
+
```python
|
|
96
|
+
from aria_trl import ContinualSFTTrainer, ARIAConfig
|
|
97
|
+
from transformers import AutoModelForSequenceClassification, AutoTokenizer, TrainingArguments
|
|
98
|
+
|
|
99
|
+
# Load model
|
|
100
|
+
model = AutoModelForSequenceClassification.from_pretrained("distilgpt2", num_labels=2)
|
|
101
|
+
tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
|
|
102
|
+
|
|
103
|
+
# Configure ARIA
|
|
104
|
+
aria_config = ARIAConfig(
|
|
105
|
+
plasticity_lambda=0.01, # bimodal gate specialization
|
|
106
|
+
spc_lambda=100.0, # Fisher consolidation strength
|
|
107
|
+
adapter_dim=64, # task adapter bottleneck
|
|
108
|
+
slow_lr_ratio=0.5, # asymmetric LR: slow=0.5x fast
|
|
109
|
+
)
|
|
110
|
+
|
|
111
|
+
# Training arguments
|
|
112
|
+
args = TrainingArguments(
|
|
113
|
+
output_dir="./checkpoints",
|
|
114
|
+
learning_rate=2e-4,
|
|
115
|
+
num_train_epochs=3,
|
|
116
|
+
per_device_train_batch_size=8,
|
|
117
|
+
)
|
|
118
|
+
|
|
119
|
+
# Create trainer
|
|
120
|
+
trainer = ContinualSFTTrainer(
|
|
121
|
+
model=model,
|
|
122
|
+
tokenizer=tokenizer,
|
|
123
|
+
args=args,
|
|
124
|
+
train_dataset=task1_train,
|
|
125
|
+
eval_dataset=task1_eval,
|
|
126
|
+
aria_config=aria_config,
|
|
127
|
+
)
|
|
128
|
+
|
|
129
|
+
# Train on task 1
|
|
130
|
+
trainer.train()
|
|
131
|
+
|
|
132
|
+
# Consolidate before next task (Fisher estimation)
|
|
133
|
+
trainer.consolidate_task(task_id=0)
|
|
134
|
+
|
|
135
|
+
# Train on task 2
|
|
136
|
+
trainer.add_task(task_id=1)
|
|
137
|
+
trainer.train_dataset = task2_train
|
|
138
|
+
trainer.eval_dataset = task2_eval
|
|
139
|
+
trainer.train()
|
|
140
|
+
|
|
141
|
+
# Consolidate again
|
|
142
|
+
trainer.consolidate_task(task_id=1)
|
|
143
|
+
|
|
144
|
+
# Evaluate on all tasks to measure forgetting
|
|
145
|
+
metrics = trainer.evaluate_all_tasks(all_tasks)
|
|
146
|
+
```
|
|
147
|
+
|
|
148
|
+
## Core Components
|
|
149
|
+
|
|
150
|
+
### ARIAConfig
|
|
151
|
+
|
|
152
|
+
Configuration for ARIA continual learning:
|
|
153
|
+
|
|
154
|
+
```python
|
|
155
|
+
from aria_trl import ARIAConfig
|
|
156
|
+
|
|
157
|
+
config = ARIAConfig(
|
|
158
|
+
plasticity_lambda=0.01, # Bimodal specialization loss weight
|
|
159
|
+
spc_lambda=100.0, # Fisher regularization strength
|
|
160
|
+
adapter_dim=64, # Task adapter bottleneck dimension
|
|
161
|
+
slow_lr_ratio=0.5, # Slow pathway LR multiplier
|
|
162
|
+
warmup_steps=500, # Before plasticity loss activates
|
|
163
|
+
consolidation_steps_per_task=None # Fisher estimation steps (None=all)
|
|
164
|
+
)
|
|
165
|
+
```
|
|
166
|
+
|
|
167
|
+
### PlasticityGatedMLP
|
|
168
|
+
|
|
169
|
+
Replaces transformer FFN layers automatically. Routes per-token computation:
|
|
170
|
+
|
|
171
|
+
```
|
|
172
|
+
output = π * fast_pathway(x) + (1−π) * slow_pathway(x)
|
|
173
|
+
```
|
|
174
|
+
|
|
175
|
+
where π ∈ (0,1) is learned and regularized to specialize (0 or 1).
|
|
176
|
+
|
|
177
|
+
### TaskFastAdapter
|
|
178
|
+
|
|
179
|
+
Per-task residual adapter (LoRA-like):
|
|
180
|
+
|
|
181
|
+
```
|
|
182
|
+
h_new = h + adapter(h)
|
|
183
|
+
```
|
|
184
|
+
|
|
185
|
+
Bottleneck design: `h → compress → ReLU → expand`, zero-initialized expansion.
|
|
186
|
+
|
|
187
|
+
### FisherConsolidator
|
|
188
|
+
|
|
189
|
+
Estimates diagonal Fisher Information on validation data after each task:
|
|
190
|
+
|
|
191
|
+
```python
|
|
192
|
+
consolidator = FisherConsolidator(model, device)
|
|
193
|
+
consolidator.consolidate(task_id, eval_loader)
|
|
194
|
+
loss = consolidator.compute_spc_loss(global_step)
|
|
195
|
+
```
|
|
196
|
+
|
|
197
|
+
## Metrics
|
|
198
|
+
|
|
199
|
+
Compute continual learning metrics with:
|
|
200
|
+
|
|
201
|
+
```python
|
|
202
|
+
from aria_trl.utils import compute_continual_metrics
|
|
203
|
+
|
|
204
|
+
metrics = compute_continual_metrics(task_accuracies)
|
|
205
|
+
# Returns: avg_accuracy, forgetting, forward_transfer
|
|
206
|
+
```
|
|
207
|
+
|
|
208
|
+
- **Average Accuracy**: Mean accuracy on all tasks at end
|
|
209
|
+
- **Forgetting**: How much old tasks degrade (lower is better)
|
|
210
|
+
- **Forward Transfer**: How much new tasks benefit from old tasks
|
|
211
|
+
|
|
212
|
+
## Example: DistilGPT2 on Sequential Tasks
|
|
213
|
+
|
|
214
|
+
See `examples/distilgpt2_example.py` for a complete working example:
|
|
215
|
+
|
|
216
|
+
```bash
|
|
217
|
+
python examples/distilgpt2_example.py
|
|
218
|
+
```
|
|
219
|
+
|
|
220
|
+
Trains on 3 sequential tasks (sentiment, toxicity, spam) and prints metrics.
|
|
221
|
+
|
|
222
|
+
## How ARIA Prevents Forgetting
|
|
223
|
+
|
|
224
|
+
### The Problem
|
|
225
|
+
Standard fine-tuning on sequential tasks overwrites weights learned on earlier tasks:
|
|
226
|
+
|
|
227
|
+
```
|
|
228
|
+
Task 1: [update all weights]
|
|
229
|
+
Task 2: [update all weights again] ← Task 1 weights overwritten
|
|
230
|
+
Task 3: [update all weights again] ← Task 1 & 2 weights overwritten
|
|
231
|
+
Eval: [accuracy on Task 1 ↓↓↓] ← Catastrophic forgetting!
|
|
232
|
+
```
|
|
233
|
+
|
|
234
|
+
### The Solution
|
|
235
|
+
ARIA separates fast (volatile) and slow (stable) pathways:
|
|
236
|
+
|
|
237
|
+
```
|
|
238
|
+
Fast pathway: [updates freely, learns task-specific patterns]
|
|
239
|
+
Slow pathway: [updates slowly, consolidated by Fisher regularization]
|
|
240
|
+
Gate π: [routes each token: fast for new patterns, slow for stability]
|
|
241
|
+
|
|
242
|
+
Result: Old tasks remain in slow pathway, new tasks learn in fast pathway
|
|
243
|
+
```
|
|
244
|
+
|
|
245
|
+
## Design Philosophy
|
|
246
|
+
|
|
247
|
+
aria-trl's design prioritizes:
|
|
248
|
+
|
|
249
|
+
1. **Compatibility**: Subclasses SFTTrainer, works with any HF model
|
|
250
|
+
2. **Simplicity**: Three mechanisms, no complex multi-model ensembles
|
|
251
|
+
3. **Efficiency**: Task adapters add <1% parameters, Fisher diagonal reduces memory
|
|
252
|
+
4. **Interpretability**: Gate π shows per-token plasticity, easy to debug
|
|
253
|
+
|
|
254
|
+
## Papers & Motivation
|
|
255
|
+
|
|
256
|
+
- **ARIA** (PyTorch): ["Adaptive Recurrent Intelligence Architecture"](https://github.com/rsd-darshan/ARIA) — core continual learning research, CNNs and task-incremental learning
|
|
257
|
+
- **aria-trl** (TRL): This package — brings ARIA mechanisms to LLM fine-tuning via Hugging Face
|
|
258
|
+
- **NCG**: ["Novelty-triggered Capacity Growth"](https://github.com/rsd-darshan/NCG) — reactive capacity expansion (predecessor)
|
|
259
|
+
- **EWC**: ["Elastic Weight Consolidation"](https://arxiv.org/abs/1612.00796) — standard baseline for catastrophic forgetting
|
|
260
|
+
|
|
261
|
+
## Limitations & Future Work
|
|
262
|
+
|
|
263
|
+
- **Slow pathway at 0.5× LR**: Conservative; tuning improves CIFAR-10 accuracy
|
|
264
|
+
- **Fisher diagonal only**: Full Fisher or Kronecker-factored approximations are future work
|
|
265
|
+
- **DistilGPT2 tested**: Larger models (7B+) need memory optimization
|
|
266
|
+
- **Binary classification**: Multi-class and language modeling tasks need validation
|
|
267
|
+
|
|
268
|
+
## Contributing
|
|
269
|
+
|
|
270
|
+
Contributions welcome! Please:
|
|
271
|
+
|
|
272
|
+
1. Fork the repo
|
|
273
|
+
2. Create a feature branch (`git checkout -b feature/my-feature`)
|
|
274
|
+
3. Commit changes (no AI attribution required)
|
|
275
|
+
4. Push to remote and open a PR
|
|
276
|
+
|
|
277
|
+
## License
|
|
278
|
+
|
|
279
|
+
MIT License — see [LICENSE](LICENSE) for details.
|
|
280
|
+
|
|
281
|
+
## Citation
|
|
282
|
+
|
|
283
|
+
If you use aria-trl in your research, please cite:
|
|
284
|
+
|
|
285
|
+
```bibtex
|
|
286
|
+
@software{poudel2026aria_trl,
|
|
287
|
+
title = {aria-trl: Continual Learning for LLM Fine-tuning},
|
|
288
|
+
author = {Poudel, Darshan},
|
|
289
|
+
year = {2026},
|
|
290
|
+
url = {https://github.com/rsd-darshan/aria-trl}
|
|
291
|
+
}
|
|
292
|
+
```
|
|
293
|
+
|
|
294
|
+
## Contact
|
|
295
|
+
|
|
296
|
+
- **GitHub**: [@rsd-darshan](https://github.com/rsd-darshan)
|
|
297
|
+
- **Email**: poudeldarshan44@gmail.com
|
|
298
|
+
|
|
299
|
+
---
|
|
300
|
+
|
|
301
|
+
**aria-trl** — continual learning, without the catastrophe.
|