Instructions to use SinhalaMMLU-SharedTask/team-sinbotics with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use SinhalaMMLU-SharedTask/team-sinbotics with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -2,7 +2,15 @@
|
|
| 2 |
|
| 3 |
## Team
|
| 4 |
- **Team name:** SinBotics
|
| 5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
## System summary
|
| 8 |
- **Base model:** [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B), revision `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a` (released 2026-02-27, Apache-2.0, open weights).
|
|
@@ -40,10 +48,6 @@ python inference.py
|
|
| 40 |
| Training validation (2,745 held-out questions from the two training sources, digit prompt) | 44.4% → 54.8% after fine-tuning |
|
| 41 |
| Final run | answers stated in the reasoning for 56.0% of questions; the rest hit the 4,096-token limit and are read off after the cue |
|
| 42 |
|
| 43 |
-
## Disclosure
|
| 44 |
-
<!-- TEAM: check and complete this section before uploading -->
|
| 45 |
-
- We chose the final configuration (base model Qwen3.5-4B, the combined-data adapter, and thinking mode) by comparing candidate systems on a 300-question subset of the shared-task test questions (100 per level), scored with answers we had for that subset. No test question or answer was used for training.
|
| 46 |
-
|
| 47 |
## Files
|
| 48 |
- `inference.py` – the exact system that produced our final Codabench submission (self-contained; see above)
|
| 49 |
- `predictions.jsonl` – the exact file we submitted to the final leaderboard
|
|
|
|
| 2 |
|
| 3 |
## Team
|
| 4 |
- **Team name:** SinBotics
|
| 5 |
+
|
| 6 |
+
**Members:**
|
| 7 |
+
|
| 8 |
+
| Name | Affiliation | E-mail |
|
| 9 |
+
|---|---|---|
|
| 10 |
+
| Sachintha Weththasinghe | SLIIT | akalankasachintha125@gmail.com |
|
| 11 |
+
| Hasitha Hirushan | SLIIT | hasithahja@gmail.com |
|
| 12 |
+
| Hasaan Saadhiq | SLIIT | hsaadhiq@gmail.com |
|
| 13 |
+
| Janith Ramanayake | UoM | janithramanayake523@gmail.com |
|
| 14 |
|
| 15 |
## System summary
|
| 16 |
- **Base model:** [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B), revision `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a` (released 2026-02-27, Apache-2.0, open weights).
|
|
|
|
| 48 |
| Training validation (2,745 held-out questions from the two training sources, digit prompt) | 44.4% → 54.8% after fine-tuning |
|
| 49 |
| Final run | answers stated in the reasoning for 56.0% of questions; the rest hit the 4,096-token limit and are read off after the cue |
|
| 50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
## Files
|
| 52 |
- `inference.py` – the exact system that produced our final Codabench submission (self-contained; see above)
|
| 53 |
- `predictions.jsonl` – the exact file we submitted to the final leaderboard
|