Listing Thumbnail

    Synthefy - Nori Fine-Tuning

     Info
    Sold by: Synthefy 
    Deployed on AWS
    Fine-tune the Synthefy-Nori tabular regression model on your own table in Parquet files with an Amazon SageMaker training job in your AWS account, then serve the fine-tuned model from an asynchronous SageMaker endpoint under a name you choose. The job starts from a released Nori base (nori-1-5, nori-6m or nori-100m), adapts it to your table and reports the validation error of the base and of the fine-tuned model on the same rows, so you can see what fine-tuning bought before you deploy.

    Overview

    Synthefy-Nori is a transformer for tabular regression that learns in context: each prediction request carries labeled rows (X_train, y_train) and the rows to predict (X_test), and the model predicts them in one forward pass without a training step. This algorithm adds a training step for a table you predict on repeatedly: it adapts a released Nori base to that table and serves the result with the same request format, so you still send fresh context rows with every prediction.

    What the job does. It reads your table from Parquet files in an S3 channel, one row group at a time, so the table can be larger than the instance's memory, trains a fine-tuned model on it with a few settings you may change (the base model, the number of models the job trains and averages, how many times each model reads its rows, the cap on training hours, the learning rate, the seed), and writes a model artifact and a training report to your bucket. The artifact is small: it holds what changed relative to the base, and the base itself stays inside the container.

    What the report tells you. Before training, the job sets aside rows that no training step sees: 10 percent of your rows up to 4,096, or the rows of your validation channel when you give one. It scores the base model and the fine-tuned model on those rows with the same context and writes both mean squared errors, val_mse_before and val_mse_after, to training_report.json, together with each model's curve, the rows read and used, the plan the job made before it trained, what ended the training and whether the job was cut short. Compare the two numbers before you deploy: fine-tuning does not improve every table.

    Use case. You predict one numeric target repeatedly from one table of about 1,400 rows or more (1,000 with a validation channel) and want a model adapted to that table while keeping Nori's in-context request format. One model can also learn several regression tasks or time series, such as one per store, from a table with a task column.

    Scope of this version. Regression only. Parquet input only: numeric columns, a finite target in every row. Trainable bases: nori-1-5 (the default) and nori-6m, on 24 GB or 48 GB GPUs, and nori-100m, on 48 GB GPUs (ml.g6e types only). One training instance per job, with 1, 4 or 8 GPUs; on several GPUs the job trains one model per GPU at a time. A job given a checkpoint folder continues after a stop or a failure from what it saved: the models it finished and the progress of a model in training. Automatic model tuning is not offered. The fine-tuned model is served from an asynchronous SageMaker endpoint and answers only to the name you set in the model_name hyperparameter, which starts with ft-.

    Where it runs. The training job, the model package and the endpoint run in your AWS account with network isolation, and your data stays in your S3 buckets. No Synthefy service is in the path, and Synthefy can neither see nor stop your jobs or endpoints. The synthefy Python package (pip install "synthefy[aws]>=9.0.0") starts and follows the job, deploys the result and calls the endpoint; plain AWS SDK or CLI calls work as well.

    Highlights

    • Fine-tunes Nori (nori-1-5, nori-6m or nori-100m) on a Parquet table, read one row group at a time, on 1, 4 or 8 GPUs. By default the job reads the whole table once; a stopped job continues from its checkpoints. The artifact holds only what changed relative to the base.
    • Every job reports val_mse_before and val_mse_after: the base and the fine-tuned model scored on the same held-out rows with the same context, so you know what fine-tuning bought before you deploy.
    • Runs in your AWS account with network isolation: the training job, model package and asynchronous endpoint are yours, and the endpoint serves the fine-tuned model under the ft- name you choose.

    Details

    Sold by

    Delivery method

    Latest version

    Deployed on AWS
    New

    Introducing multi-product solutions

    You can now purchase comprehensive solutions tailored to use cases and industries.

    Multi-product solutions

    Features and programs

    Financing for AWS Marketplace purchases

    AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
    Financing for AWS Marketplace purchases

    Pricing

    Synthefy - Nori Fine-Tuning

     Info
    This product is available free of charge. Free subscriptions have no end date and may be canceled any time.
    Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator  to estimate your infrastructure costs.

    AI Insights

     Info

    Dimensions summary

    This product carries no software charge; you pay only for the SageMaker instances you run. Dimensions split into two activities: training and inference, each billed per host hour. Training runs on several instance sizes across the g5, g6, and g6e families, from xlarge up to 48xlarge, so you pick capacity to match your fine-tuning job. Inference bills per host hour on g5 and g6 xlarge and 2xlarge sizes. Inference also separates into real-time mode, for live requests, and batch mode, for bulk jobs. You scale cost by choosing instance size and run time.

    Top-of-mind questions for buyers

    You pay for each hour one instance stays running. One host hour means one instance active for one hour. If you run several instances at once, each one meters its own hours. Charges start when the endpoint reaches a running state and stop when it stops.
    Host-hour charges apply while an instance runs. A stopped or deleted endpoint stops accruing software charges. Keeping a real-time endpoint active to serve live requests meters every hour it stays up, even between requests. Delete the endpoint when idle to stop charges.
    Real-time inference keeps an endpoint running to serve live requests, so you pay for every hour it stays active. Batch inference runs as a bulk job, so you pay only for the hours the job runs, then stops. Batch suits large one-time prediction sets.
    docs.synthefy.com
    Helpful?

    Vendor refund policy

    Synthefy-Nori Fine-Tuning is offered free of charge on AWS Marketplace, so there are no software fees to refund. Buyers pay only their own AWS infrastructure costs (for example, SageMaker training and endpoint instance charges), which are billed directly by AWS and are outside Synthefy's control. For any questions, contact support@synthefy.com .

    How can we make this page better?

    Tell us how we can improve this page, or report an issue with this product.
    Tell us how we can improve this page, or report an issue with this product.

    Legal

    Vendor terms and conditions

    Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .

    Content disclaimer

    Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.

    Usage information

     Info

    Delivery details

    Amazon SageMaker algorithm

    An Amazon SageMaker algorithm is a machine learning model that requires your training data to make predictions. Use the included training algorithm to generate your unique model artifact. Then deploy the model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.

    Deploy the model on Amazon SageMaker AI using the following options:
    Before deploying the model, train it with your data using the algorithm training process. You're billed for software and SageMaker infrastructure costs only during training. Duration depends on the algorithm, instance type, and training data size. When training completes, the model artifacts save to your Amazon S3 bucket. These artifacts load into the model when you deploy for real-time inference or batch processing. For more information, see Use an Algorithm to Run a Training Job  .
    Deploy the model as an API endpoint for your applications. When you send data to the endpoint, SageMaker processes it and returns results by API response. The endpoint runs continuously until you delete it. You're billed for software and SageMaker infrastructure costs while the endpoint runs. AWS Marketplace models don't support Amazon SageMaker Asynchronous Inference. For more information, see Deploy models for real-time inference  .
    Deploy the model to process batches of data stored in Amazon Simple Storage Service (Amazon S3). SageMaker runs the job, processes your data, and returns results to Amazon S3. When complete, SageMaker stops the model. You're billed for software and SageMaker infrastructure costs only during the batch job. Duration depends on your model, instance type, and dataset size. AWS Marketplace models don't support Amazon SageMaker Asynchronous Inference. For more information, see Batch transform for inference with Amazon SageMaker AI  .
    Version release notes

    Whole-table fine-tuning with Nori 1.5 as a base, with synthefy 9.0.0.

    Training

    • nori-1-5 is a base model and the default; nori-6m and nori-100m remain.
    • The hyperparameters say how much of the table each model reads: n_passes_per_row (unset: every row once) or n_steps_per_model, row_partition_strategy (shard or folds) and max_training_hours (24 by default, at most 119.5) take the place of steps and budget_s, which a job of this version refuses.
    • patience counts validation scores in a row without a better one (25 by default; none: every step of the plan), and validation_every_n_steps sets the training steps between two scores.
    • A job of tasks fine-tunes one model on several regression tasks or time series: task_column, time_column, context_rows, query_rows and tasks.
    • Checkpoints: a job given a checkpoint folder (CheckpointConfig) saves each model it finishes and the progress of a model in training. A job with resume set to true continues from the folder; a new job needs an empty folder.
    • The job logs its plan (models, rows, steps and hours) before it trains. The report adds training_plan, stop_reason, n_passes_per_row_completed, n_rows_unread, unstarted_children, fraction_of_rows_as_targets, seconds_per_step, bag and row_partition_strategy.

    Serving

    • Unchanged: the asynchronous endpoint serves the fine-tuned model under its ft- name, with the request and response of version 1.1.

    Instance types

    • Training: ml.g5.xlarge, ml.g5.2xlarge, ml.g6.xlarge, ml.g5.12xlarge, ml.g5.48xlarge (nori-1-5 and nori-6m); ml.g6e.xlarge, ml.g6e.2xlarge, ml.g6e.12xlarge, ml.g6e.48xlarge (all three bases).
    • Endpoint: ml.g5.xlarge, ml.g5.2xlarge, ml.g6.xlarge.

    Client

    • synthefy 9.0.0 or later. Version 1.1 keeps steps and budget_s for synthefy 8.x.

    Additional details

    Inputs

    Summary

    The fine-tuned model keeps Nori's in-context request format: each request carries labeled context rows and the rows to predict, and the response has one prediction per query row.

    Content type: application/json (SynthefyNoriClient sends the same request as application/vnd.synthefy.nori.tensorfile).

    Request body (one JSON object):

    • model: required string, the fine-tuned model's name, equal to the model_name hyperparameter of the training job (ft-...). Any other value is refused.
    • X_train: required array of arrays, the labeled context rows, shape [n_context][n_features]. Numbers or null.
    • y_train: required array of numbers, one finite target per X_train row, aligned by position.
    • X_test: required array of arrays, the rows to predict, with the same columns in the same order as X_train.
    • task: optional string, "regression" (the default and only value).

    Use the training table's feature columns, in training order, without the target, encoded the same way as for training.

    Asynchronous endpoint: the request is an object in your S3 staging prefix passed to InvokeEndpointAsync. Batch transform: one request per S3 object (SplitType None).

    Limitations for input type
    Regression only. Features must be numeric (missing values may be null); encode categorical columns before sending. X_train and X_test must have the same number of columns in the same order. SageMaker Asynchronous Inference accepts an input object up to 1 GB and processes an invocation for up to one hour; practical table size depends on GPU and host memory. The client does not split oversized requests.
    Input MIME type
    application/json
    {"model":"ft-marketplace-validation","task":"regression","X_train":[[0.12,-1.05,0.33,0.80,-0.41,1.20,0.05,-0.66],[-0.73,0.48,-1.12,0.27,0.90,-0.35,1.41,0.18],[1.05,0.22,0.61,-0.94,-1.30,0.44,-0.27,0.73],[-0.38,-0.59,0.95,1.16,0.21,-0.88,0.62,-1.02]],"y_train":[0.31,0.87,-0.42,1.64],"X_test":[[0.40,-0.20,0.10,0.55,-0.75,0.30,0.95,-0.15],[-1.10,0.85,-0.45,-0.60,0.35,1.05,-0.50,0.25]]}
    {"model":"ft-marketplace-validation","task":"regression","X_train":[[0.12,-1.05,0.33,0.80,-0.41,1.20,0.05,-0.66],[-0.73,0.48,-1.12,0.27,0.90,-0.35,1.41,0.18],[1.05,0.22,0.61,-0.94,-1.30,0.44,-0.27,0.73],[-0.38,-0.59,0.95,1.16,0.21,-0.88,0.62,-1.02]],"y_train":[0.31,0.87,-0.42,1.64],"X_test":[[0.40,-0.20,0.10,0.55,-0.75,0.30,0.95,-0.15],[-1.10,0.85,-0.45,-0.60,0.35,1.05,-0.50,0.25]]}

    Input data descriptions

    The following table describes supported input data fields for real-time inference and batch transform.

    Field name
    Description
    Constraints
    Required
    model
    Name of the fine-tuned model, the training job's model_name
    ft- then 1 to 60 lowercase letters, digits and -, the first a letter or digit
    Yes
    X_train
    Labeled context rows, shape [n_context][n_features]
    Numbers or null; same columns as X_test
    Yes
    y_train
    Regression target of each X_train row
    Finite numbers; one per X_train row
    Yes
    X_test
    Rows to predict, shape [n_query][n_features]
    Same columns and order as X_train
    Yes
    task
    Task to run
    "regression" only; default "regression"
    No

    Support

    Vendor support

    Support is available by email at support@synthefy.com . We assist with subscription, training jobs, deployment, invocation, and functional questions on a best-effort basis, typically responding within two business days. Training jobs and endpoints run in your AWS account, so include the training job name and its FailureReason, or the endpoint name, when you write.

    AWS infrastructure support

    AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.

    Similar products

    Customer reviews

    Ratings and reviews

     Info
    0 ratings
    5 star
    4 star
    3 star
    2 star
    1 star
    0%
    0%
    0%
    0%
    0%
    0 reviews
    No customer reviews yet
    Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.