Skip to main content

Fixed: "Incompatible input types for column … Can not safely convert int64 to float64" in MLflow

· 7 min read
Jagdish Kumawat
Founder @ Dewiride

My model refused its own kind of data: a CSV file of shopping sessions failed with Incompatible input types for column Administrative. Can not safely convert int64 to float64. The model's signature declares those columns as decimals, double, and pandas reads whole numbers from a CSV as 64-bit integers. MLflow converts several types on its own, but not 64-bit integers to doubles. Reading the signature's double columns as float64 fixed my own code; for MLflow's command line, sending JSON instead of CSV did.

What I Was Doing​

I was building the predict command of the shopper model in Part 6 of the MLflow series, on MLflow 3.18.0 and pandas 3.1.1. The model was trained with its whole-number columns read as decimals, so its signature lists them as double. The command reads a CSV of sessions, checks it against the signature and predicts. Its first version read the file plainly:

src/shopper_purchase_prediction/predict.py (first version, excerpt)
sessions = pd.read_csv(args.input)
model = mlflow.pyfunc.load_model(args.model_uri)
try:
# The signature lists every column the model needs and its type. Checking first turns
# a bad file into one clear sentence instead of a failed prediction.
mlflow.models.validate_schema(sessions, model.metadata.get_input_schema())

The file was three sessions I made up, with whole numbers where the dataset has counts and codes:

examples/sessions.csv
Administrative,Administrative_Duration,Informational,Informational_Duration,ProductRelated,ProductRelated_Duration,BounceRates,ExitRates,PageValues,SpecialDay,Month,OperatingSystems,Browser,Region,TrafficType,VisitorType,Weekend
3,80.0,1,30.0,40,1500.0,0.0,0.01,25.0,0.0,Nov,2,2,1,2,Returning_Visitor,False
0,0.0,0,0.0,1,0.0,0.2,0.2,0.0,0.0,May,1,1,1,3,New_Visitor,True
1,20.0,0,0.0,12,300.0,0.01,0.03,0.0,0.0,Mar,2,2,3,1,Returning_Visitor,False
Terminal
uv run predict --model-uri models:/m-a56a76fad37f4c4e886eaf465f57e322 --input examples/sessions.csv

What I Expected​

A purchase probability for each of the three sessions.

What I Got​

Output
2026/11/21 13:45:05 ERROR shopper_purchase_prediction: examples/sessions.csv doesn't match the model's signature: Incompatible input types for column Administrative. Can not safely convert int64 to float64.

The predict command's output in a terminal: an ERROR saying examples/sessions.csv doesn't match the model's signature: Incompatible input types for column Administrative. Can not safely convert int64 to float64.

The same file, scored with MLflow's own command line and a later model from the same training, failed the same way:

Terminal
uv run mlflow --env-file .env models predict -m models:/m-6e2a8bb07fc44b02843eca07c555586f -i examples/sessions.csv -t csv --env-manager local
Output (the end of the last line)
Error: Incompatible input types for column Administrative. Can not safely convert int64 to float64.

The end of the mlflow models predict output for the CSV file in a terminal: the error Incompatible input types for column Administrative, Can not safely convert int64 to float64.

What I Checked​

MLflow's guide to signatures lists the conversions it makes on its own: "int → long # 32-bit to 64-bit integer", "int → double # Integer to float", "float → double". And among the rejected ones: "long → double # Potential precision loss".

What pandas made of the file. It read Administrative and the other whole-number columns as int64, a 64-bit integer: MLflow's "long".

MLflow's code. The check in mlflow/models/utils.py allows integers to become doubles only when they're at most 6 bytes wide, with one exception: a 64-bit column holding nothing but 0, 1 or missing values. That exception is why a column of zeros passed in an earlier test of mine, and why the real data didn't.

Why It Happens​

A double has 53 bits of precision (Python's floating-point guide: "IEEE 754 binary64 values contain 53 bits of precision"), while a 64-bit integer uses 64, so some large integers can't be represented exactly as doubles. MLflow therefore refuses to convert 64-bit integers to doubles. A CSV file doesn't say which columns are decimals, so pandas picks int64 for any column of whole numbers, and the signature's double columns get int64 data. JSON input doesn't have this problem: MLflow parses it with the model's signature in hand.

The Fix​

Read every column the signature declares as double as float64, taking the list from the model itself:

The fixed lines of predict.py in a macOS-style code window: after loading the model, it gets the input schema, builds a decimals mapping of every double column to float64 with a comment explaining why, reads the CSV with dtype set to that mapping, and then validates the schema.

src/shopper_purchase_prediction/predict.py (excerpt)
model = mlflow.pyfunc.load_model(args.model_uri)
schema = model.metadata.get_input_schema()
# A CSV file doesn't say which numbers are decimals, so pandas reads whole numbers as
# 64-bit integers, which MLflow won't turn into the doubles the signature asks for.
# Read those columns as decimals from the start.
decimals = {
column.name: "float64" for column in schema.inputs if column.type == DataType.double
}
sessions = pd.read_csv(args.input, dtype=decimals)

DataType comes from mlflow.types. For MLflow's command line, which reads the CSV itself, send the rows as JSON instead, in the dataframe_split layout of the model's serving_input_example.json.

After the Fix​

The command, writing its results to a file this time:

Terminal
uv run predict --model-uri models:/m-6e2a8bb07fc44b02843eca07c555586f --input examples/sessions.csv --output scored.csv
Output
2026/11/21 15:00:03 INFO shopper_purchase_prediction: Wrote 3 scored sessions to scored.csv

The predict command after the fix in a terminal: Wrote 3 scored sessions to scored.csv.

And the command line, with the same three sessions as JSON, sessions.json:

Terminal
uv run mlflow --env-file .env models predict -m models:/m-6e2a8bb07fc44b02843eca07c555586f -i sessions.json -t json --env-manager local
Output (last line)
{"predictions": [[0.18206545331405966, 0.8179345466859403], [0.9994173242880812, 0.0005826757119188792], [0.99863978394165, 0.0013602160583500364]]}

The mlflow models predict command with the JSON file in a terminal: the probabilities for the three sessions, 0.8179, 0.0006 and 0.0014 of buying.

The JSON held the same whole numbers, and passed.

Common Mistakes​

MistakeWhat happensDo this instead
Training on integer columns to avoid the errorThe signature says long, and a missing value is refused insteadKeep double, and read input as float64
Testing the conversion with a column of zerosIt passes, by MLflow's 0-or-1 exception, and hides the problemTest with real values

Troubleshooting​

SymptomLikely causeFix
Can not safely convert int64 to float64 from mlflow models predict -t csvThe CLI reads the CSV without the signature's typesSend JSON with -t json
The same error from your own codepd.read_csv without dtypeRead the signature's double columns as float64
Can not safely convert float64 to int64The opposite: decimals for a long columnMake the column whole numbers, or log the model with double

FAQ​

Why does MLflow call it long?​

In MLflow's types, int is a 32-bit integer and long a 64-bit one; its guide: "int → long # 32-bit to 64-bit integer". pandas' int64 is 64-bit, so it's a long.

Does the model care whether the numbers are 3 or 3.0?​

No: the model was trained on these columns as decimals, and 3.0 is exactly 3.

Conclusion​

The signature asked for double and pandas gave int64, a conversion MLflow refuses for loss of precision. Reading the signature's double columns as float64 let my command score the file, and JSON input let MLflow's command line do the same.

Additional Resources​