README.md

---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
library_name: transformers
---

# Granite 3.0 8B Instruct - Uncertainty LoRA

<< @Connor Leech Some creative, fun copy should go here describing how this is being shared as an IBM Research experiment that we want to put out in the world and get some early feedback.  Try it out and let us know how it goes (but don't be upset if it breaks / isn't in our products just yet)!>>


## Model Summary

**Granite 3.0 8b Instruct - Uncertainty** is a LoRA adapter for [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct), 
adding the capability to provide calibrated certainty scores when answering questions when prompted, in addition to retaining the full abilities of the [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct) model. 

- **Developer:** IBM Research
- **Model type:** LoRA adapter for [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct)
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)


### Model Sources

<!-- Provide the basic links for the model. -->


- **Paper:** The **Granite Uncertainty 3.0 8b** model is finetuned to provide certainty scores mimicking the output of a calibrator trained via the method in [[Shen et al. ICML 2024] Thermometer: Towards Universal Calibration for Large Language Models](https://arxiv.org/abs/2403.08819)


## Usage

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

### Intended use

<<@Connor to rewrite this statement as needed>> This is an experimental LoRA testing new functionality being developeed for IBM's Granite LLM family.  We are welcoming the community to test it out and give us feedback, but we are NOT recommending this model be used for real deployments at this time.  Stay tuned for more updates on the Granite roadmap.

**Granite Uncertainty 3.0 8b** is lightly tuned so that its behavior closely mimics that of [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct), 
with the added ability to generate certainty scores for answers to questions when prompted. 

**Certainty score definition** The model will respond with a certainty percentage, quantized to 10 possible values (i.e. 5%, 15%, 25%,...95%). 
This percentage is *calibrated* in the following sense: given a set of answers assigned a certainty score of X%, approximately X% of these answers should be correct. See the eval experiment below for out-of-distribution verification of this behavior.

**Important note** Certainty is inherently an intrinsic property of a model and its abilitities. **Granite Uncertainty 3.0 8b** is not intended to predict the certainty of responses generated by any other models besides itself or [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct).

**Usage steps** Answering a question and obtaining a certainty score proceeds as follows. 

1. Prompt the model with a system prompt (required) followed by the user prompt. The model is calibrated to work best with the system prompt provided below. 
2. Use the model to generate a response as normal (via the `assistant` role).
3. Prompt the model to generate a certainty score by generating in the `certainty` role (use "certainty" as the role in the chat template, or simply append `<|start_of_role|>certainty<|end_of_role|>` and continue generating), see examples below.
4. The model will respond with a certainty percentage, quantized with steps of 10% (i.e. 05%, 15%, 25%,...95%).  Note, any additional text after the score and % can be ignored.  You can curb additional generation by setting "max token length" = 3 when using this role.

When not given the certainty generation prompt `<|start_of_role|>certainty<|end_of_role|>`, the model's behavior should mimic that of the base model [ibm-granite/granite-3.0-8b-instruct](https://huggingface.co/ibm-granite/granite-3.0-8b-instruct). 

**System prompt** The model was calibrated with the following system prompt: `You are an AI language model developed by IBM Research. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.` 
You can further augment this system prompts for a given use case or task, but it is recommended your system prompt always starts with this string.



### Quickstart Example

The following code describes how to use the Granite Uncertainty model to answer questions and obtain intrinsic calibrated certainty scores. Note that a generic system prompt is included, this is not necessary and can be modified as needed.

```python
import torch,os
from transformers import AutoTokenizer,  AutoModelForCausalLM
from peft import PeftModel, PeftConfig

token = os.getenv("HF_MISTRAL_TOKEN")
BASE_NAME = "ibm-granite/granite-3.0-8b-instruct"
LORA_NAME = "ibm-granite/granite-uncertainty-3.0-8b-lora"
device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Load model
token = os.getenv("HF_MISTRAL_TOKEN")
tokenizer = AutoTokenizer.from_pretrained(BASE_NAME,padding_side='left',trust_remote_code=True, token=token)
model_base = AutoModelForCausalLM.from_pretrained(BASE_NAME,device_map="auto")
model_UQ = PeftModel.from_pretrained(model_base, LORA_NAME)

system_prompt = "You are an AI language model developed by IBM Research. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior." 
question = "What is IBM?"
print("Question:" + question)
question_chat = [
	{
		"role": "system",
		"content": system_prompt
	},
	{
		"role": "user",
		"content": question
	},
]

# Generate answer
input_text = tokenizer.apply_chat_template(question_chat,tokenize=False,add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt")
output = model_UQ.generate(inputs["input_ids"].to(device), attention_mask=inputs["attention_mask"].to(device), max_new_tokens=80)
output_text = tokenizer.decode(output[0])
answer = output_text.split("assistant<|end_of_role|>")[1]
print("Answer: " + answer)

# Generate certainty score
uq_generation_prompt = "<|start_of_role|>certainty<|end_of_role|>"
uq_chat = [
    {
        "role": "system",
        "content": system_prompt
    },
    {
        "role": "user",
        "content": question
    },
    {
        "role": "assistant",
        "content": answer
    },
]

uq_text = tokenizer.apply_chat_template(uq_chat,tokenize=False) + uq_generation_prompt
inputs = tokenizer(uq_text, return_tensors="pt")
output = model_UQ.generate(inputs["input_ids"].to(device), attention_mask=inputs["attention_mask"].to(device), max_new_tokens=1)
output_text = tokenizer.decode(output[0])
uq_score = int(output_text[-1])
print("Certainty: " + str(5 + uq_score * 10) + "%")
```


## Evaluation

The model was evaluated on the [MMLU](https://huggingface.co/datasets/cais/mmlu) datasets (not used in training). Shown are the [Expected Calibration Error (ECE)](https://towardsdatascience.com/expected-calibration-error-ece-a-step-by-step-visual-explanation-with-python-code-c3e9aa12937d) for each task, for the base model (Granite-3.0-8b-instruct) and Granite-Uncertainty-3.0-8b.
The average ECE across tasks for our method is 0.064 (out of 1) and is consistently low across tasks (maximum task ECE 0.10), compared to the base model average ECE of 0.20 and maximum task ECE of 0.60. Note that our ECE of 0.064 is smaller than the gap between the quantized certainty outputs (10% quantization steps). Additionally, the zero-shot performance on the MMLU tasks does not degrade, averaging at 89%.
<!-- This section describes the evaluation protocols and provides the results. -->



![image/png](/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F6602ffd971410cf02bf42c06%2F2MwP7DRZlNBtWSKWFvXOI.png%3C%2Fspan%3E)



## Training Details
The **Granite Uncertainty 3.0 8b** model is a LoRA adapter finetuned to provide certainty scores mimicking the output of a calibrator trained via the method in [[Shen et al. ICML 2024] Thermometer: Towards Universal Calibration for Large Language Models](https://arxiv.org/abs/2403.08819).



### Training Data
The following datasets were used for calibration and/or finetuning. 

* [BigBench](https://huggingface.co/datasets/tasksource/bigbench)
* [MRQA](https://huggingface.co/datasets/mrqa-workshop/mrqa)
* [newsqa](https://huggingface.co/datasets/lucadiliello/newsqa)
* [trivia_qa](https://huggingface.co/datasets/mandarjoshi/trivia_qa)
* [search_qa](https://huggingface.co/datasets/lucadiliello/searchqa)
* [openbookqa](https://huggingface.co/datasets/allenai/openbookqa)
* [web_questions](https://huggingface.co/datasets/Stanford/web_questions)
* [smiles-qa](https://huggingface.co/datasets/alxfgh/ChEMBL_Drug_Instruction_Tuning)
* [orca-math](https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k)
* [ARC-Easy](https://huggingface.co/datasets/allenai/ai2_arc)
* [commonsense_qa](https://huggingface.co/datasets/tau/commonsense_qa)
* [social_i_qa](https://huggingface.co/datasets/allenai/social_i_qa)
* [super_glue](https://huggingface.co/datasets/aps/super_glue)
* [figqa](https://huggingface.co/datasets/nightingal3/fig-qa)
* [riddle_sense](https://huggingface.co/datasets/INK-USC/riddle_sense)
* [ag_news](https://huggingface.co/datasets/fancyzhx/ag_news)
* [medmcqa](https://huggingface.co/datasets/openlifescienceai/medmcqa)
* [dream](https://huggingface.co/datasets/dataset-org/dream)
* [codah](https://huggingface.co/datasets/jaredfern/codah)
* [piqa](https://huggingface.co/datasets/ybisk/piqa)



## Model Card Authors 

Kristjan Greenewald