Spaces:
Sleeping
Sleeping
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: green
|
| 6 |
sdk: docker
|
| 7 |
pinned: false
|
| 8 |
-
short_description:
|
| 9 |
hf_oauth: true
|
| 10 |
hf_oauth_expiration_minutes: 36000
|
| 11 |
hf_oauth_scopes:
|
|
@@ -16,14 +16,40 @@ hf_oauth_scopes:
|
|
| 16 |
- read-billing
|
| 17 |
tags:
|
| 18 |
- autotrain
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
---
|
| 20 |
|
| 21 |
-
#
|
| 22 |
|
| 23 |
-
|
| 24 |
|
| 25 |
-
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
@misc{thakur2024autotrainnocodetrainingstateoftheart,
|
| 28 |
title={AutoTrain: No-code training for state-of-the-art models},
|
| 29 |
author={Abhishek Thakur},
|
|
@@ -31,5 +57,5 @@ https://huggingface.co/docs/autotrain
|
|
| 31 |
eprint={2410.15735},
|
| 32 |
archivePrefix={arXiv},
|
| 33 |
primaryClass={cs.AI},
|
| 34 |
-
url={https://arxiv.org/abs/2410.15735},
|
| 35 |
-
}
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Gemma-3-4B-PT Full-Model Reasoning Research
|
| 3 |
+
emoji: 🧠
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: green
|
| 6 |
sdk: docker
|
| 7 |
pinned: false
|
| 8 |
+
short_description: Researching multimodal SFT logic on Gemma-3-4B-PT
|
| 9 |
hf_oauth: true
|
| 10 |
hf_oauth_expiration_minutes: 36000
|
| 11 |
hf_oauth_scopes:
|
|
|
|
| 16 |
- read-billing
|
| 17 |
tags:
|
| 18 |
- autotrain
|
| 19 |
+
- gemma
|
| 20 |
+
- multimodal
|
| 21 |
+
- reasoning
|
| 22 |
+
- sft
|
| 23 |
---
|
| 24 |
|
| 25 |
+
# 🎯 Project Objective: Improving Multimodal Logic in Gemma 3
|
| 26 |
|
| 27 |
+
This Space is dedicated to an educational research project focused on **Full-Model Supervised Fine-Tuning (SFT)** of the `google/gemma-3-4b-pt` architecture.
|
| 28 |
|
| 29 |
+
The goal is to move beyond standard Low-Rank Adaptation (LoRA) to observe how full-parameter updates affect the model's ability to handle complex chain-of-thought reasoning across multimodal inputs.
|
| 30 |
|
| 31 |
+
## 🛠️ Hardware Requirements & Grant Justification
|
| 32 |
+
* **Baseline:** Nvidia A10G-small (24GB VRAM)
|
| 33 |
+
* **Preferred:** **Nvidia A10G-large** (Additional CPU RAM for sharding/preprocessing)
|
| 34 |
+
|
| 35 |
+
Because Gemma 3 is a multimodal model, the vision-language alignment layers and the full-parameter gradient states require the 24GB VRAM capacity of the A10G. Using an A10G-large will allow for faster dataset tokenization and more efficient model sharding during the "Push to Hub" phase, reducing the total grant time used.
|
| 36 |
+
|
| 37 |
+
## 🧪 Methodology
|
| 38 |
+
- **Training Type:** Full-Model SFT (Supervised Fine-Tuning)
|
| 39 |
+
- **Precision:** `bf16` with `adamw_bnb_8bit` optimizer
|
| 40 |
+
- **Data:** Curated reasoning dataset formatted in ChatML for logical consistency.
|
| 41 |
+
|
| 42 |
+
## 🤝 Community Commitment
|
| 43 |
+
As per the grant request, once training is finalized:
|
| 44 |
+
1. The **full model weights** will be pushed to the Hub.
|
| 45 |
+
2. Training logs (Loss curves/Perplexity) will be made public.
|
| 46 |
+
3. **The Space will be manually reverted to the Free CPU tier to release resources back to the community.**
|
| 47 |
+
|
| 48 |
+
# 📜 Docs & Citation
|
| 49 |
+
|
| 50 |
+
Official Documentation: [AutoTrain Docs](https://huggingface.co/docs/autotrain)
|
| 51 |
+
|
| 52 |
+
```bibtex
|
| 53 |
@misc{thakur2024autotrainnocodetrainingstateoftheart,
|
| 54 |
title={AutoTrain: No-code training for state-of-the-art models},
|
| 55 |
author={Abhishek Thakur},
|
|
|
|
| 57 |
eprint={2410.15735},
|
| 58 |
archivePrefix={arXiv},
|
| 59 |
primaryClass={cs.AI},
|
| 60 |
+
url={[https://arxiv.org/abs/2410.15735](https://arxiv.org/abs/2410.15735)},
|
| 61 |
+
}
|