StructuredLLM-7b.GGUF

StructuredLLM models aim to map a target data object consisting of a title and corresponding example values to a set of input titles and their example values. Mapping can be one-to-one or one-to-many.

Model Details

Model Description

• Model type: LLM
• Finetuned from model [optional]: mistralai/Mistral-7B-v0.1

Inference

.gguf models can be inferenced using llama.cpp (llama-cpp-python). Please follow the instructions within these repos to get started.

Inference Examples

This is an instruct finetune utilizing the Alpaca instruct format (introduced by stanford-alpaca):

"{system_prompt}\n\n### Instruction:\n{instruction}\n\n### Response: "

The model is finetuned to perform the following task:

Structured Data Mapping

Expected input:

You are a business assistant that specialized in normalizing JSON data into a consistent structure.
You will be given tasks such as assessing which keys in an Input-JSON object map to a given Target-JSON.
The usecase of the information you respond with will be to carry over and transform data in the direction Input-JSON -> Target-JSON.

Keys can be considered valid mappings when either their names indicate the same or a very similar concept, or when their values are a close match in what they represent (salary, dates, IDs, etc.) or have similar formats. Mappings do not need to be an exact match, only have sufficient overlap.
For some cases, multiple keys from the Input-JSON might be required to map to the Target-JSON. This is the case if the value in the Target-JSON can only be arrived at due to the information present under multiple Input-JSON keys.

### Instruction:
Map the following Input-JSON to the given Target-JSON:

Input-JSON:
{
    "FormEntryDate_INT": [
        "1615785600",
        "1615910400",
        "1616131200"
    ],
    "SPECIAL INDEX 2": [
        "#00010",
        "#00321",
        "#00543"
    ],
    "RET_BEN_VEST_Y5": [
        "Fifty thousand, one dollar and eleven cents",
        "Ninety-nine thousand, nine hundred ninety-nine dollars and ninety-nine cents",
        "One thousand, two hundred thirty-four dollars and fifty-six cents"
    ],
    "COM_GOV_GRD": [
        "R",
        "S",
        "T"
    ],
    "FSA_Trns_Elct_Sts": [
        "ELE",
        "STAT",
        "WAIV"
    ],
    "Variable Pay Structure ID": [
        "VP-018/str",
        "VP-019/str",
        "VP-020/str"
    ],
    "ParkingSlotID": [
        "P10018",
        "P10019",
        "P10020"
    ],
    "Ds3.DdQr": [
        "2,718",
        "4,057",
        "3,951"
    ],
    "CARRIER DEFINED REPORTING 6": [
        "CR6Definition",
        "R6ByCarrier",
        "CarrierDefinition6"
    ],
    "EE MaritalStatus": [
        "Married but Reserved",
        "Living Life as a Bachelor",
        "Can't Be Tamed"
    ],
    "PredictionMatrix_2030": [
        "1.025",
        "1.242",
        "0.832"
    ],
    "PensionFund[9]": [
        "PRM-CHC-FND",
        "ELITE-PEN-SCH",
        "PLAT-PEN-PLAN"
    ],
    "TOTAL_AMT": [
        "$125.75",
        "$790.80",
        "$975.65"
    ],
    "TrendRate": [
        "1.0",
        "1.0",
        "1.0"
    ],
    "YearlyValueChange": [
        "0.01",
        "0.04",
        "0.01"
    ],
    "EeEducationLevel": [
        "BS Diploma",
        "MA",
        "MS Degree"
    ],
    "retnedEarnsPrcntg": [
        "0.90 of Total Earnings",
        "0.95 of Total Earnings",
        "1.00 of Total Earnings"
    ]
}

Target-JSON:
{
    "PredictionMatrix_2040": [
        "0.325",
        "0.918",
        "0.752"
    ]
}

### Response: 

Expected Output:

{
    "reasoning": "1. The target key 'PredictionMatrix_2040' likely refers to a prediction matrix for the year 2040. 2. The values of the target data are floating point numbers, indicating a certain rate or ratio. 3. The 'PredictionMatrix_2030' input key provides similar floating point values, which could be used to calculate or estimate the 'PredictionMatrix_2040' target key. The 'TrendRate' and 'YearlyValueChange' input keys could also potentially factor into calculating the target key's value.",
    "mapped_input_keys": [
        "PredictionMatrix_2030",
        "TrendRate",
        "YearlyValueChange"
    ]
}

Training Data

zero-systems/ColumnMapping.10k.INSTRUCT

Training Methodology

StructuredLLM was trained using QLoRA. Resulting adapter was merged into the base model weights, converted to the gguf format and finally quantized to 4 bits.