The Generation Result is dedicated in the file Generation.json
. File contains:
-
model_id: the reference id for the generative model
-
prompt_template: the template for the prompt for generate the result
-
result:
- reference_index: paper index given in the dataset
- DOI: DOI reference for the paper
- generation: generated result
The Ground Truth is dedicated in the file Ground_Truth.json
. File contains:
- reference_index: paper index given in the dataset
- DOI: DOI reference for the paper
- ground_truth: annotated result
The Generation range for each item is given below:
"Light_source": ["UV", "Solar", "UV-Vis", "Monochromatic", "Solar Simulator"],
"Lamp" : ["Fluorescent", "Mercury", "Halogen", "Mercury-Xenon", "LED", "Tungsten", "Xenon", "Tungsten-Halide", "Solar Simulator"],
"Reactor_type": ["Slurry", "Fixed-bed", "Optical Fiber", "Monolithic", "Membrane", "Fluidised-bed"],
"Reaction_medium": ["Liquid", "Gas"],
"Operation_mode" : ["Batch", "Continuous", "Batch/Continuous"]
Result based on LLama-3-70B:
generation:
{
"catalyst": " TiO2",
"co_catalyst": " Ag",
"light_source": " UV",
"lamp": " Hg",
"reaction_medium": " Liquid",
"reactor_type": " Slurry",
"operation_mode": " Batch"
}
The Average accuarcy for each item, calculated according to Evaluation Process
in README
.
The evaluation result is dedicated in Evaluation.json
. File contains:
-
generation_model_id: id reference for the generation model
-
similarity_model_id: id refernce for the similarity model
-
source_ground_truth: path for the file that contains the ground_truth
-
source_generation: path for the file that contains the generation result
-
evaluation_strategy: the evaluation strategy we adopt, detailed in
Evaluation Process
inREADME
-
metric: the evaluation metric
-
result:
- item: the targeted item
- value: evaluation numerical value based on the evalution metric
Result based on LLama-3-70B:
evaluation:
{
"generation_model_id": "meta-llama/Meta-Llama-3-70B-Instruct",
"similarity_model_id": "Salesforce/SFR-Embedding-Mistral",
"source_ground_truth": "/Solar/result/LLama_3_70B/Ground_Truth.json",
"source_generation": "/Solar/result/LLama_3_70B/Generation.json",
"evaluation_strategy": "rule-based",
"metric": "accuracy",
"result": [
{"item": "catalyst",
"acc": 0.8275862068965517},
...
]
}
The context or chunks that RAG system has selected to provide the context for the generative model.
The context is dedicated in Context.json
. File contains:
-
similarity_model_id: id refernce for the similarity model
-
similarity_method: the method of calculating similarity
-
context:
-
reference_index: paper index given in the dataset
-
contexts:
- item: targeted item
- context: a list of all the selected chunks from the original paper
-
context:
{
"similarity_model_id": "Salesforce/SFR-Embedding-Mistral",
"similarity_method": "Cosine_Similarity",
"context": [
{"reference_index": "1",
"context": {
"item": ["Operation_mode"],
"chunk": ["XXXXX", "XXXXX"]
}},
...
]
}
SolarFactRAG is a RAG approach that based on the extracted facts from chunks. The process of extracting facts from chunks are given:
- Use LLM to extract entities from the chunks
- Based on the extracted entities and the corresponding chunks, use llm to extract relations between entities.
- Use LLM to constructing simple, short factual sentences based on entities and relations.
Here is the table that describe the saved output json file:
Category | Definition | DataType | Example |
---|---|---|---|
paper_title | The title of the paper | String | Effect of silver doping on the TiO2 for photocatalytic reduction of CO2 |
DOI | The DOI of the paper(extracted from the paper pdf file) | String | 10.1016/j.apcatb.2010.02.030 |
generation_model | The ollama model id for llm generation | String | llama3.2:3b |
similarity_model | The ollama model id for text embedding | String | nomic-embed-text |
similarity_metric | The metric for calculating the similarity between embeddings | String | Cosine_Similarity |
result | The list of generation result from the llm | List | Details given velow |
The details for result
part of the saved output json file:
Category | Definition | DataType | Example | Range(if applicable) |
---|---|---|---|---|
question_category | The category of the question for the llm | String | Light_source | [catalyst, co_catalyst, Light_source, Lamp, Reaction_medium, Reactor_type, Operation_mode] |
query | The query for the llm | String | What is the Light_source used in the experiment? | Not applicable |
generation | The generated answer from the llm | String | ###Light_source: UV | Not applicable |
RAG_source | The information source provided for the RAG pipeline | String | generated_facts | Not applicable |
selected_answer | The corresponding answer from the selection of choices | String | UV | Details are given below |
evidences | The list of evidences for the RAG pipeline | list | Details given below | Not applicable |
The range for selected_answer
with correspongind question_category
:
question_category | Range(if applicable) |
---|---|
catalyst | Not applicable |
co_catalyst | Not applicable |
Light_source | 'UV', 'Solar', 'UV-Vis', 'Monochromatic', 'Solar Simulator', 'Do not Know' |
Lamp | 'Fluorescent', 'Mercury', 'Halogen', 'Mercury-Xenon', 'LED', 'Tungsten', 'Xenon', 'Tungsten-Halide', 'Solar Simulator', 'Do not Know' |
Reaction_medium | 'Liquid', 'Gas', 'Do not Know' |
Reactor_type | 'Slurry', 'Fixed-bed', 'Optical Fiber', 'Monolithic', 'Membrane', 'Fluidised-bed', 'Do not Know' |
Operation_mode | 'Batch', 'Continuous', 'Batch/Continuous', 'Do not Know' |
The details for evidence
part of the result
from the output json:
Category | Definition | DataType | Example |
---|---|---|---|
similarity_score | The similarity score between the query and correspond text provided to the embedding model | Float | 0.6205 |
pdf_reference | The original text that are extracted from the paper | String | in the conventional focusing... |
generated_fact | The generated facts based on the pdf_reference | String | Facts: 1. A homemade apparatus is used... |