Update README.md
Browse files
README.md
CHANGED
|
@@ -121,6 +121,8 @@ The model was trained using the [ServiceNow/BigDocs-Sketch2Flow](https://hugging
|
|
| 121 |
* Trained with **16× NVIDIA H100 80GB GPUs** across two nodes.
|
| 122 |
* Full Sharded Data Parallel (FSDP) training, no CPU offloading.
|
| 123 |
|
|
|
|
|
|
|
| 124 |
## Evaluation
|
| 125 |
|
| 126 |
### Testing Data
|
|
@@ -156,11 +158,15 @@ All Evaluation metrics can be found in the official [StarFlow repo](https://gith
|
|
| 156 |
|
| 157 |
Finetuning yields **large gains over base Pixtral-12B and GPT-4o**, particularly in matching workflow components and triggers.
|
| 158 |
|
|
|
|
|
|
|
| 159 |
## Model Examination
|
| 160 |
|
| 161 |
* Finetuned models capture **naming conventions** and structured execution logic better.
|
| 162 |
* Failure modes include **missing ELSE branches** or **generic table names**.
|
| 163 |
|
|
|
|
|
|
|
| 164 |
## Technical Specifications
|
| 165 |
|
| 166 |
### Model Architecture and Objective
|
|
@@ -173,6 +179,8 @@ Finetuning yields **large gains over base Pixtral-12B and GPT-4o**, particularly
|
|
| 173 |
* **Hardware:** 16× NVIDIA H100 80GB (2 nodes)
|
| 174 |
* **Software:** FSDP, bf16 mixed precision, PyTorch/Transformers
|
| 175 |
|
|
|
|
|
|
|
| 176 |
## Citation
|
| 177 |
|
| 178 |
**BibTeX:**
|
|
@@ -189,6 +197,8 @@ Finetuning yields **large gains over base Pixtral-12B and GPT-4o**, particularly
|
|
| 189 |
**APA:**
|
| 190 |
Béchard, P., Wang, C., Abaskohi, A., Rodriguez, J., Pal, C., Vazquez, D., Gella, S., Rajeswar, S., & Taslakian, P. (2025). **StarFlow: Generating Structured Workflow Outputs from Sketch Images**. *arXiv preprint arXiv:2503.21889*.
|
| 191 |
|
|
|
|
|
|
|
| 192 |
## Glossary
|
| 193 |
|
| 194 |
* **FlowSim**: Metric based on tree edit distance for workflows.
|
|
@@ -196,16 +206,22 @@ Béchard, P., Wang, C., Abaskohi, A., Rodriguez, J., Pal, C., Vazquez, D., Gella
|
|
| 196 |
* **Trigger Match**: Correctness of predicted workflow trigger.
|
| 197 |
* **Component Match**: Correctness of predicted components (order-agnostic).
|
| 198 |
|
|
|
|
|
|
|
| 199 |
## More Information
|
| 200 |
|
| 201 |
* [ServiceNow Flow Designer](https://www.servicenow.com/products/platform-flow-designer.html)
|
| 202 |
-
*
|
|
|
|
|
|
|
| 203 |
|
| 204 |
## The StarFlow Team
|
| 205 |
|
| 206 |
* Patrice Béchard, Chao Wang, Amirhossein Abaskohi, Juan Rodriguez, Christopher Pal, David Vazquez, Spandana Gella, Sai Rajeswar, Perouz Taslakian
|
| 207 |
|
|
|
|
|
|
|
| 208 |
## Model Card Contact
|
| 209 |
|
| 210 |
* Patrice Bechard - [patrice.bechard@servicenow.com](mailto:patrice.bechard@servicenow.com)
|
| 211 |
-
* ServiceNow Research – [research.servicenow.com](https://research.servicenow.com)
|
|
|
|
| 121 |
* Trained with **16× NVIDIA H100 80GB GPUs** across two nodes.
|
| 122 |
* Full Sharded Data Parallel (FSDP) training, no CPU offloading.
|
| 123 |
|
| 124 |
+
---
|
| 125 |
+
|
| 126 |
## Evaluation
|
| 127 |
|
| 128 |
### Testing Data
|
|
|
|
| 158 |
|
| 159 |
Finetuning yields **large gains over base Pixtral-12B and GPT-4o**, particularly in matching workflow components and triggers.
|
| 160 |
|
| 161 |
+
---
|
| 162 |
+
|
| 163 |
## Model Examination
|
| 164 |
|
| 165 |
* Finetuned models capture **naming conventions** and structured execution logic better.
|
| 166 |
* Failure modes include **missing ELSE branches** or **generic table names**.
|
| 167 |
|
| 168 |
+
---
|
| 169 |
+
|
| 170 |
## Technical Specifications
|
| 171 |
|
| 172 |
### Model Architecture and Objective
|
|
|
|
| 179 |
* **Hardware:** 16× NVIDIA H100 80GB (2 nodes)
|
| 180 |
* **Software:** FSDP, bf16 mixed precision, PyTorch/Transformers
|
| 181 |
|
| 182 |
+
---
|
| 183 |
+
|
| 184 |
## Citation
|
| 185 |
|
| 186 |
**BibTeX:**
|
|
|
|
| 197 |
**APA:**
|
| 198 |
Béchard, P., Wang, C., Abaskohi, A., Rodriguez, J., Pal, C., Vazquez, D., Gella, S., Rajeswar, S., & Taslakian, P. (2025). **StarFlow: Generating Structured Workflow Outputs from Sketch Images**. *arXiv preprint arXiv:2503.21889*.
|
| 199 |
|
| 200 |
+
---
|
| 201 |
+
|
| 202 |
## Glossary
|
| 203 |
|
| 204 |
* **FlowSim**: Metric based on tree edit distance for workflows.
|
|
|
|
| 206 |
* **Trigger Match**: Correctness of predicted workflow trigger.
|
| 207 |
* **Component Match**: Correctness of predicted components (order-agnostic).
|
| 208 |
|
| 209 |
+
---
|
| 210 |
+
|
| 211 |
## More Information
|
| 212 |
|
| 213 |
* [ServiceNow Flow Designer](https://www.servicenow.com/products/platform-flow-designer.html)
|
| 214 |
+
* [StarFlow Blog](https://www.servicenow.com/blogs/2025/starflow-ai-turns-sketches-into-workflows)
|
| 215 |
+
|
| 216 |
+
---
|
| 217 |
|
| 218 |
## The StarFlow Team
|
| 219 |
|
| 220 |
* Patrice Béchard, Chao Wang, Amirhossein Abaskohi, Juan Rodriguez, Christopher Pal, David Vazquez, Spandana Gella, Sai Rajeswar, Perouz Taslakian
|
| 221 |
|
| 222 |
+
---
|
| 223 |
+
|
| 224 |
## Model Card Contact
|
| 225 |
|
| 226 |
* Patrice Bechard - [patrice.bechard@servicenow.com](mailto:patrice.bechard@servicenow.com)
|
| 227 |
+
* ServiceNow Research – [research.servicenow.com](https://research.servicenow.com)
|