|
--- |
|
datasets: |
|
- SALT-NLP/positive_reframing |
|
language: |
|
- en |
|
license: bigscience-bloom-rail-1.0 |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# Model Card for Model ID |
|
|
|
This model is a BLOOM-base adjusted to the sentiment transfer task, developed as part of a FourthBrain workshop on GenerativeAI. |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
This model is a BLOOM-base adjusted to the sentiment transfer task, where the objective is to reverse the sentiment polarity of a text without contradicting |
|
the original meaning. Positive reframing induces a complementary positive viewpoint (e.g. glass-half-full) escaping negative patterns. |
|
Based on the article [Ziems at. al (2022)](https://arxiv.org/abs/2204.02952). |
|
|
|
Sample working space [here](https://huggingface.co/spaces/telmo000/bloom-positive-reframing). |
|
|
|
### Input |
|
`### Negative sentence:\n{original_text}\n\n### Reframing strategy: \n{reframing_strategy}\n\n### Reframing sentence:\n` |
|
|
|
|
|
- **Developed by:** Telmo Correa |
|
- **Model type:** LLM |
|
- **Language(s) (NLP):** English |
|
- **License:** [bigscience-bloom-rail-1.0](https://bigscience.huggingface.co/blog/the-bigscience-rail-license) |
|
- **Finetuned from model :** [bigscience/bloom-1b7](https://huggingface.co/bigscience/bloom-1b7) |
|
|
|
## Uses |
|
|
|
Model trained as a proof-of-concept fine tuning of BLOOM for sentence rewriting. |
|
|
|
### Direct Use |
|
|
|
Model is intended to be directly used to rewrite sentences with the provided strategy. |
|
|
|
### Out-of-Scope Use |
|
|
|
Any uses of the model must abide by the terms of both the original BLOOM model and the Salt-NLP/positive-reframing dataset. |
|
|
|
## Bias, Risks, and Limitations |
|
|
|
As a fine-tuned version of BLOOM, this model carries all the biases, risks, and limitations. of its original training. |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
[Salt-NLP/positive-reframing](https://huggingface.co/datasets/SALT-NLP/positive_reframing) |
|
|
|
### Training Procedure |
|
|
|
The baseline model [bigscience/bloom-1b7](https://huggingface.co/bigscience/bloom-1b7) was trained through 100 steps over the training split of the training data, with its prompt engineered to request explicit positive sentence reframing: |
|
|
|
``` |
|
Below is a negative sentence, please select a reframing strategy and write the positive reframed sentence. |
|
|
|
### Negative sentence: |
|
NEGATIVE SENTENCE HERE |
|
|
|
### Reframing strategy: |
|
STRATEGY HERE |
|
|
|
### Reframed sentence: |
|
REFRAMED SENTENCE HERE |
|
``` |
|
|
|
|
|
#### Training Hyperparameters |
|
|
|
- **Training regime:** fp16 non-mixed precision, using PEFT and LoRA |
|
|
|
## Evaluation |
|
|
|
Evaluation not performed. |
|
|
|
## Environmental Impact |
|
|
|
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly --> |
|
|
|
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). |
|
|
|
- **Hardware Type:** Coogle Colab PRO GPU |
|
- **Hours used:** 10 min |
|
- **Cloud Provider:** GCP |
|
- **Compute Region:** us-west-1 |
|
- **Carbon Emitted:** 10g |
