pronics2004 mnagired commited on
Commit
784e5e4
·
1 Parent(s): 1ee3f6e

Update factuality-correction/granite-4.0-micro/README.md (#6)

Browse files

- Update factuality-correction/granite-4.0-micro/README.md (e79b7a13ff8f4dbf33845eb0d8e0c7fe378c9ffd)


Co-authored-by: Manish Nagireddy <mnagired@users.noreply.huggingface.co>

factuality-correction/granite-4.0-micro/README.md CHANGED
@@ -154,9 +154,11 @@ If you find this detector useful, please cite the following work.
154
  }
155
  ```
156
 
157
- **Infrastructure:**
 
158
 
159
- **Ethical Considerations:** This model is part of an ongoing research effort focused on post-generation mitigation and remains experimental and under active development. We are committed to continuous improvement and welcome constructive feedback to enhance its performance and capabilities. It is also important to note that there is no built-in safeguard to guarantee that the corrected response will always be factually correct. As with other generative models, safety assurance relies on offline evaluations (see [Evaluation](#evaluation)), and we expect, but cannot ensure, that the corrected response meets safety standards.
 
160
 
161
  **Resources**
162
 
 
154
  }
155
  ```
156
 
157
+ ## Infrastructure
158
+ We trained the Granite Micro 4.0 Factuality Correction adapter on IBM's Vela cluster using 4 A100 GPUs.
159
 
160
+ ## Ethical Considerations
161
+ The Granite Micro 4.0 Factuality Correction adapter is primarily fine-tuned on English-only input–output pairs. Although the underlying base model supports multilingual dialogue, the adapter’s performance on non-English tasks may differ from its performance on English. In addition, while the base model has been aligned with safety considerations in mind, the adapter may, in some cases, produce inaccurate, biased, or otherwise unsafe outputs in response to user prompts. It is also important to note that there is no built-in safeguard guaranteeing that the correction output is always correct. As with other generative models, safety assurance relies on offline evaluation procedures (see Evaluation), and while we expect the generated outputs to meet safety standards, this cannot be guaranteed. Finally, this adapter is specifically optimized for the factuality definition described above, and its behavior outside that scope may be limited.
162
 
163
  **Resources**
164