LLM-Blender

• LLM-Blender is an ensembling framework designed to achieve consistently superior performance by combining the outputs of multiple language models (LLMs). This work focuses on integrating LLM-Blender with Retrieval-Augmented Generation (RAG) pipelines to significantly improve the quality of generated text.

• LLM-Blender is a two-stage ensemble learning framework. In the first stage (ranking), pairwise comparison of candidates is performed, and they are then ranked. In the second stage (fusing), the top K candidates are merged to render the final output.

• The LLM-Blender comprises of two modules: the PairRanker and the GenFuser. The PairRanker module compares the outputs from multiple LLMs to provide the top-ranked outputs. It compares each candidate with the input in a pairwise manner, making it robust to subtle differences in the generated text. The GenFuser module uses the top-ranked outputs from the PairRanker module to generate an improved output. The module fuses the top K of the N-ranked candidates from the PairRanker, conditioned on the input instruction, to generate an enhanced output.

• A custom Haystack component, LLMBlenderRanker, has been implemented to integrate LLM-Blender with Haystack pipelines. The component utilizes the PairRanker module from the LLM-Blender framework, which compares each candidate with the input in a pairwise manner. Different LLMs can generate subtly different texts, since they are trained on different datasets and tasks. By comparing each text in a pairwise manner, the component ranks and ensembles the text so it is robust to these subtle differences.

• Haystack RAG Pipelines with the LLM-Blender component to ensemble LLMs were evaluated. The pipelines were evaluated on the BillSum and MixInstruct datasets using three metrics: BARTScore, BLEURT, and BERTScore. The llama-3, phi-3, mistral-7b, openchat-3.5, starling-lm-7b-alpha and openhermes-2.5 LLMs were used in the ensemble.

PairRanker

• The PairRanker module is responsible for comparing and ranking the outputs from LLM's. During the ranking stage, a specific input prompt (x) is passed to N different LLMs, and their outputs are compiled as candidates ($y_1$, …, $y_N$).

• The PairRanker then analyzes and ranks these candidates. For each input x, the candidates are obtained from N different LLMs. This input sequence, along with the candidates, is then subjected to a cross-attention text encoder, such as RoBERTa. The text encoder is tasked with learning and determining the superior candidate for the given input x.

• All the candidates are paired ($y_i$ and $y_j$), producing a matrix of pairwise comparison results. These pairs are evaluated based on the condition: given the input prompt, which candidate's output is better? By aggregating the results in the matrix, the PairRanker can rank all candidates and take the top K of them for generative fusion.

GenFuser

• The primary goal of the GenFuser module is to capitalize on the strengths of the top K selected candidates from the PairRanker's ranking.

• After the PairRanker module ranks the candidates, the GenFuser module is employed to fuse the top K out of the N ranked candidates and generate an improved final output. It takes a seq2seq approach, fusing the set of top candidates while conditioning on the input prompt, to generate an improved and enhanced output.

RAG Pipeline with the LLM Blender component

The results from the different LLMs on the MixInstruct dataset are ranked and combined using the LLM-Blender framework.

Usage

To run the pipelines, you will need to clone this repository and install the required libraries. Install the llm-blender package:

git clone https://github.com/avnlp/llm_blender
cd llm_blender
pip install -e .

LLM-Blender using Mistral, LLama-3 and Phi-3 models on the MixInstruct Dataset

cd src/llm_blender/mix_instruct/
python llm_blender_ranker_all_llms.py

LLMBlenderRanker Component Usage

llm_ranker = LLMBlenderRanker(model="llm-blender/PairRM")
answers = [
    GeneratedAnswer(data="Paris is the capital of France.", query="What makes Paris unique?", documents=[]),
    GeneratedAnswer(
        data="The Eiffel Tower is an iconic landmark in Paris.", query="What makes Paris unique?", documents=[]
    ),
    GeneratedAnswer(data="Berlin is a beautiful city.", query="What makes Paris unique?", documents=[]),
]
output = llm_ranker.run(answers=answers)
ranked_answers = output["answers"]
print(ranked_answers)

# [
#     GeneratedAnswer(
#         data="The Eiffel Tower is an iconic landmark in Paris.",
#         query="What makes Paris unique?",
#         documents=[],
#         meta={},
#     ),
#     GeneratedAnswer(
#         data="Paris is the capital of France.", query="What makes Paris unique?", documents=[], meta={}
#     ),
#     GeneratedAnswer(data="Berlin is a beautiful city.", query="What makes Paris unique?", documents=[], meta={}),
# ]

The API documentation can be found here.

Results

• A custom component, LLMBlenderPairRanker, was developed to integrate the LLM-Blender Framework with Haystack Pipelines. Haystack RAG Pipelines with the LLM-Blender component to ensemble LLMs were evaluated. The pipelines were evaluated on the BillSum and MixInstruct datasets using three metrics: BARTScore, BLEURT, and BERTScore.

• We successfully replicated the previously reported results for the LLM-Blender. Moreover, significantly improved performance was observed when utilizing newer LLM models, such as Llama-3-8B, Phi-3-mini and Mistral-7B. These findings demonstrate the potential of ensembling state-of-the-art LLMs to enhance the performance of RAG Pipelines on question-answering, summarization and instruction-following tasks.

• The authors of LLM-Blender obtained BERTScore values in the range of 62.26 to 74.68 on the MixInstruct dataset. They obtained a BERTScore value of 72.97 with the PairRanker. We obtained BERTScore values in the range of 72.62 to 76.86 using the newer LLMs. We obtained a BERTScore value of 75.83 with the PairRanker ensembling the results from Llama-3-8B, Phi-3-mini and Mistral-7B.

• The authors of LLM-Blender obtained BARTScore values in the range of -4.57 to -3.14 on the MixInstruct dataset. They obtained a BARTScore value of -3.14 with the PairRanker. We obtained BARTScore values in the range of -3.17 to -2.87 using the newer LLMs. We obtained a BARTScore value of -2.87 with the PairRanker ensembling the results from Llama-3-8B, Phi-3-mini and Mistral-7B.

• The authors of LLM-Blender obtained BLEURT values in the range of -1.23 to -0.37 on the MixInstruct dataset. They obtained a BLEURT value of -0.37 with the PairRanker. We obtained BLEURT values in the range of -0.41 to -0.23 using the newer LLMs. We obtained a BLEURT value of -0.26 with the PairRanker ensembling the results from Llama-3-8B, Phi-3-mini and Mistral-7B.

• The newer models like Llama-3-8B, Phi-3-mini, and Mistral-7B significantly outperformed all the models used by the LLM Blender authors on all the three metrics: BERTScore, BARTScore and BLEURT on the MixInstruct dataset.

• On the BillSum dataset, we obtained BERTScore values from 73.91 to 75.43, BARTScore values from -3.49 to -3.19, and BLEURT values from -0.39 to -0.20 across the different LLMs. The PairRanker model, ensembling the outputs from Llama-3-8B, Phi-3-mini, and Mistral-7B, achieved the highest scores of 75.83 for BERTScore, -3.19 for BARTScore, and -0.20 for BLEURT.

• For both the BillSum and MixInstruct datasets, the PairRanker model achieved the best performance when ensembling the outputs from Llama-3-8B, Phi-3-mini, and Mistral-7B. This combination of LLMs, ensembled using the LLM Blender, significantly outperformed each individual model's performance on all the evaluation metrics.

License

The source files are distributed under the MIT License.