BIGST_BAA_Topic 01_GFM.docx

UNCLASSIFIED

BIG-ST BAA HM047623BAA0001

Topic 01: GFM_ 12/15/2023

UNCLASSIFIED

BIG-ST BAA HM047623BAA0001

Topic 01: GFM_ 12/15/2023

BIG-ST BAA (HM0476-23-BAA-0001)

Topic # 01 – Geospatial-Intelligence Foundation Model (GFM)

For instructions on submission of abstracts and proposals and evaluation of proposals, please review BIG-ST BAA General Solicitation (HM047623BAA0001).

Document Change History

Title Geospatial-Intelligence Foundational Model (GFM) Expected Awards The Government plans to make one award, but may make multiple awards or none.

The amount of resources made available under this BAA Topic will depend on the quality of the proposals received, the availability of funds, and the proven value of the capabilities developed in subsequent years.

Budget The estimated range for the base and option periods:

Topic Technology Readiness Level (TRL): 2-6 This BAA Topic is limited to projects that meet Technology Readiness Level (TRL) definitions in the TRL range 2-6. If the base effort develops a promising technology demonstration (TRL 2-3), budget will be allocated in Option Year (OY) 1 for architecture optimization (TRL 3-4), in OY2 for application design and component integration (TRL 4-5), and in OY3 for integration and testing (TRL 5-6).

Our aim in the base period is to develop a trained foundational neural network model able to ingest overhead and ground-level imagery and vector data. The design solution will answer geospatial questions with accuracy and precision while demonstrating a proof of concept sufficient to seek further development in this area.

Dates Table 1. Dates

Point of Contact (POC) Please send all questions and submissions to the Contracting Officer (CO): Daniela Garavito Email Address: BIGSTBAA@nga.mil ➔ Use the following email subject for all email correspondence: ‘TOPIC 01 –GFM’

Topic Research Opportunity Description This research topic will explore approaches to developing a new type of Multimodal Foundation Model (MFM) to ingest overhead and ground-level imagery, vector data, terrain data, and ground level images, then show adaptability to perform novel, previously unencountered tasks without being retrained. It will answer specific geospatial questions with verifiable precision and accuracy paired with a rigorously determined level of confidence that can be expressed to an analyst, which will communicate the amount of trust that can be placed in the machine generated answer. It will demonstrate how performance scales with the size of the network designed for geospatial information.

This research topic covers GFM, geospatial artificial intelligence (GeoAI), multimodal machine learning, contrastive machine learning, masked autoencoding, unsupervised machine learning, semi- and self-supervised machine learning, fully supervised machine learning, transformer architectures, cross view image matching, neural networks, Vision Transformers, Shifted Windows (SWIN) Transformers, remote sensing, and geographic information system (GIS).

The goal of this research will be to deliver a geographically aware foundational model--or assembly of foundation models--that can form components of a powerful virtual assistant for analysts to geolocate media or objects or answer questions requiring specific geospatial understanding. By the end of the program, the assistant should be able to save analysts significant time in searching for, and within, geospatial sources. Both the questions and responses provided will not be linguistic; instead they will be graphically or pictorially expressed. Successful demonstrations based on this GFM will include:

· Geolocating ground level imagery using image content alone to within 1000m of ground truth.

· Geolocating aerial imagery or a collection of point line and area geospatial features without assisting geo-registered metadata to within 500m of ground truth.

· Reconstructing correctly labeled and attributed point, line, and area geospatial features from remotely sensed imagery (generating a map where map features have attributes), and generating plausible overhead and ground level imagery from point line and area based geo-registered data input.

· Appropriately transforming between panchromatic and multispectral overhead imagery, generating seasonally correct image predictions.

Background and Purpose In 2022, some significant new capabilities were released to the public based on large “foundation” machine learning models. These foundation models were Generative Pre-training Transformer (GPT) and Contrastive Language Image Pretraining (CLIP). The following information is provided to demonstrate our understanding of existing capabilities and NGA’s desire to seek novel solutions that are not solely dependent on previously explored capabilities.

GPT is a Large Language Model (LLM) of 1.7 trillion parameters designed to efficiently predict sequences of words based on its training using trillions of “tokens,” encoding text from millions of internet-based sources, and having the emergent capability of being able to answer questions interactively[footnoteRef:1]. GPT is capable of answering verbally posed questions with eloquent verbal answers; this technology forms the foundation of the popular ChatGPT capability that is now commonly used to write poetry and novel stories from simple prompts or to summarize or expand on text it is given in longer engineered prompts. [1: https://openai.com/research/gpt-4]

CLIP is a Large Machine Learning Model of 63 million parameters trained with a “contrastive” technique using 400 million image-language pairs encoding bits of text with related tokens encoding bits of photos. Based on the “embeddings” (correlations) created by the contrastive training between text and photos CLIP is able to generate accurate captions and descriptions of image prompts[footnoteRef:2]. When paired with a stable diffusion model, CLIP can generate realistic images based on text prompts. CLIP is the technology foundation for the popular Dall-E3 capability that is now commonly used to create graphics and digital art[footnoteRef:3]. [2: https://arxiv.org/pdf/2111.09734.pdf] [3: https://openai.com/dall-e-2]

More recently, in April 2023, Scale-aware Masked Auto Encoding (ScaleMAE) was published[footnoteRef:4]. This is the new state of the art in reconstructing geo-registered imagery. NGA is now embarking on a DoD Frontier project to combine the ScaleMAE technique with a Large Language Model to construct an automated image interpretation assistant[footnoteRef:5]. [4: https://arxiv.org/abs/2212.14532] [5: https://hpc.mil/solution-areas/resource-management/frontier-projects]

An older model that NGA has performed some experimentation with is GeoEstimation[footnoteRef:6]. This 2018 model pioneers one method of incorporating geospatial location in the training data to geolocate photos. [6: https://openaccess.thecvf.com/content_ECCV_2018/papers/Eric_Muller-Budack_Geolocation_Estimation_of_ECCV_2018_paper.pdf ]

NGA experimentation with these models has highlighted some challenges to their use in our field. First, they tend to “hallucinate” and have a significant rate of producing incorrect answers because of linguistic word substitutions that fail to properly use precise language. In the case below, the use of the preposition “near” is problematic because the coordinates may be near on a global scale, but perhaps would not be considered near for an analyst. Large Language Models are not trained with geospatial data, so they are exceedingly poor at answering geospatial questions accurately or asking for clarification on tasks.

The statement regarding coordinates not being able to identify a precise place is incorrect and indicates that the linguistic model alone is the wrong tool to use to answer geospatial questions. Similar errors are seen using CLIP based models, because the CLIP model embeddings correlate images and words, but not specific geolocations, thus any question requiring complex or specific geospatial understanding will not be answered correctly by such a foundation model alone.

While ScaleMAE shows wonderful capability at being able to reconstruct satellite imagery at multiple levels of resolution and it is trained with some location information, it does not include robust geospatial feature understanding or language data. Any linguistically posed questions requiring geospatial preposition understanding, geospatial distance understanding, or cardinal direction understanding as those that might be posed by an analyst will be beyond its unassisted capability to answer.

GeoEstimation, partly because of its specific focus on crossview image geolocalization and because of its size, is not a sufficiently robust general geospatial Earth model to support a variety of geospatial queries. Its current accuracy for performing even the task for which it was designed is below the threshold required for NGA.

The insufficiencies of geospatial services built using machine learning approaches are well documented in a paper by Gengchen Mai, et al. “On the Opportunities and Challenges of Foundation Models for Geospatial Intelligence”[footnoteRef:7]. There exists no machine learning based capability to derive accurate answers from a variety of information in heterogeneous geospatial data types, as is illustrated below, or to combine several machine learning based capabilities to assemble a virtual team of assistants to accomplish this task. [7: https://arxiv.org/abs/2304.06798]

Project Structure/Scope The GFM project will address the insufficiencies of current foundational models by producing a foundation model with a heterogeneous set of geospatial data to support a digital assistant capable of answering intelligence questions requiring geospatial understanding. In simple terms, it will “teach” the machine what the Earth is like all over so it can answer questions accurately and communicate its confidence in those answers.

The research requested will explore architectures for correlating multiple types of geospatial data. This approach to correlated data will improve accuracy of responses to geospatially related questions from the analyst.

3.1 Base Period – Proof of Concept

The base period is for a scientific investigative effort (TRL 2-3). This period will explore various architectures to determine those most likely to be successful. The goal of the base period is to adapt neural network architectures and principles to the field of geospatial reasoning so that the answers produced are accurate, precise, and high confidence. At the conclusion of the base period, an industry leading proof of concept GFM shall be created and demonstrated. This base period research will also demonstrate any scaling relationship between the size of the network (number of parameters) and its performance. Neural network architectures were invented more than 50 years ago; however, only recently was it realized that extending their scale into the trillion-parameter range for language data would result in new emergent behavior. In the event that a larger scale model for geospatial data is indicated by this research, then additional resources may be applied.

The fundamental challenges that need to be addressed are:

Challenge 1 – Network Design:

Engineer an appropriate network with a sufficient number of parameters to be able to contain the required information. The network architecture design shall be based on theory guided by the task and data to be processed. This will involve several iterations at different scales to determine how scaling the network improves performance. This network architecture and design needs to be complete, implemented, and initially trained sufficiently to demonstrate industry leading capability, even if scaling to a much larger model is indicated by the research.

Challenge 2 – Tokenizing Method:

Develop a method for encoding geo-registered points lines and complex polygons into “tokens” so that the information contained in the data can be captured by the network for training and subsequent reasoning. For the purposes of this project, text and picture tokenization methods have been sufficiently explored and can be adopted from publications, if needed. Tokenization methods for pictures will be required in this research because images are an important component of the training data. New tokenization methods will need to be invented for encoding the shape and attributes stored in geospatial points, lines, and complex polygons. The type of geospatial data shown below is typical of what will need to be used for training.

In the upper left, this graphic shows remote sensing images from different dates that can be used as training data to understand what an area looks like from above in different seasons. It shows six different geospatial data sets that roughly correlate with the two remotely sensed images to inform the network about the geospatial features that correlate with the remotely sensed imagery. Finally, it also shows two ground level images of the thousands that exist in that area that illustrate how those features appear from a ground level perspective. Other data sets like population density data, terrain, or elevation data may also be important to use for training; however, in this particular example the terrain is flat.

Geospatial information tokenization methods must be designed to work with the neural network or transformer-based network designed under Challenge 1. The tokenization method must be designed, implemented, and used to ingest Open Street Map data. This geospatial data will be correlated with ground level imagery, terrain data, and remotely sensed imagery in the architecture specified under Challenge 1. The cross correlation of these data sets will form a set of geographically unique “embeddings” across the Earth, so that specific geospatial questions can be answered given inputs well aligned with any of the input data types.

Optional Challenge 3 – ML Teaming:

Investigate teaming of machine learning systems. While not required to investigate this as part of this effort, NGA will recognize and support work in this area if performers describe how this type of architecture will lead to efficiency gains over a single model.

Optional Challenge 4 - Machine-to-Machine Interfaces or Continuous Learning:

Investigate interfaces with non-machine learning based systems or continual learning systems. While not required to investigate this as part of this effort, NGA will recognize and support work in this area if performers describe how they will mitigate the issue of “concept drift” through a hybrid architecture more complex than a simple machine learning approach.

Most major machine learning based models in use have been trained with data sets that were only current up to the time the model began to be used. Because the world continues to evolve past the time when these models were trained, they “drift” in their understanding of the present reality. There are three techniques to mitigate this issue. One is to make systems continuously learn[footnoteRef:8] so that the “concept drift” does not occur[footnoteRef:9]. Another is to engineer a system capable of accessing up-to-date databases or more rigorous computational systems. This is how online AI’s like Google Bard and Microsoft Bing operate, by allowing them to query the open internet prior to providing an answer to a query. To our knowledge, there does not exist an interface between an AI system and a large GIS. The final approach to solve this problem is through prompt engineering as is now frequently done with LLMs. [8: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8259323/] [9: https://www.ibm.com/docs/en/spss-modeler/18.3.0?topic=nuggets-continuous-machine-learning]

3.1.4.1 Option Period 1 - Optimization

During Option Period 1, the performer will refine the best performing proof of concept developed in the base period into a laboratory proven capability that can be delivered to NGA Research for independent experimentation and testing, (TRL 3-4). The capability is not expected to be sufficiently developed for actual analytical use, but should clearly demonstrate how effectively a trained model of the type designed will perform if constructed.

3.1,4.2 Option Period 2 – Component Integration During Option Period 2, the performer shall refine the capability for NGA’s integration into a Government-owned environment (TRL 4-5), which will then be tested independently by the performer in their facility. Afterwards, NGA will conduct testing with relevant classified subjects. By the end of this option period, an initial fully automated, machine learning trained, digital assistant for analysts shall be demonstrated by the performer.

3.1.4.3 Option Period 3- Relevant Environment Testing

In the previous option period, NGA will have developed an initially integrated digital assistant making use of the GFM capability. The four challenges (if applicable) will have been addressed and may only need refinement. That refinement of the capability will result in regular software deliveries for independent testing by NGA both in the unclassified and classified environments. During Option Period 3, the performer will further refine the code base and complete integration with other Machine Learning assemblies. Other assemblies will be discussed as part of development after contract award. (TRL 5-6)

3.1.4.4 Outside the Scope of this Topic

· Solutions employing only large language models or CLIP type models. The research has to address geospatial point line and polygon data and relationships between multiple types of geo-registered data.

· An approach where GIS data or geospatial vectors are transformed into a rasterized format and encoded for the network like an overhead image will not be an approach considered in this research because this approach will not be able to encode the attributes associated with the geospatial data.

· Any original data collection efforts.

· Any significant hardware purchases, unless justified as lowest cost.

Performance Metrics Performers shall meet the following metrics:

Table 2. Technical Requirements –Metrics and Milestones Table

The size of the area of uncertainty associated with each geolocation.

(smaller is better).

The confidence associated with the distance and uncertainty area size (between 0 and 1).

The size of the area of uncertainty associated with each geolocation (smaller is better).

The confidence associated with the distance and uncertainty area size (between 0 and 1).

Value as judged by NGA analytical evaluators on a scale of 0 to 100.

The size of the area of uncertainty associated with each geolocation (smaller is better).

To measure progress throughout the course of the project, some specific metrics must be used. For a project involving a machine learning capability with graphical output, not all of these metrics can be quantitatively evaluated so the value of progress, especially in later phases, may be judged by the qualitative utility of the evolving system to the analysts. The following tests are being considered for the resultant model or system of models:

4.1 Photograph to Location: When presented with a photograph, predict within 1000m the geographic location in which it was captured. This will require correlation between ground images, satellite images, terrain data, a geographic reference system, and worldwide geospatial data coverage to constrain the search. NGA maintains testing data with over 100,000 geolocated images worldwide to be used for this. The ground truth geolocations of the images will not be released for this BAA. The effectiveness of the tool as measured by this metric will be how close the derived geolocations are to the ground truth, the area of uncertainty associated with each geolocation, and how well the confidence estimate represents its true performance. Evaluation for this measure is quantitative and repeatable. Successful efforts must far surpass state of the art as defined by https://www.researchgate.net/publication/360311029_Where_in_the_World_is_this_Image_Transformer-based_Geo-localization_in_the_Wild

4.2 `Satellite Image to Location: When presented with a satellite image, whether panchromatic or multispectral, predict within 500m the geographic location in which it was captured regardless of the season in which it was taken. This will require training with satellite images and worldwide geospatial data coverage to constrain the search and a method for indexing global satellite imagery. NGA will easily be able to have a test set of thousands of satellite image chips of various resolutions and spectral attributes to test this capability. The ground truth geolocations of the images will not be released as part of the BAA. The effectiveness of the tool as measured by this metric will be how close the derived geolocations are to the ground truth, the area of uncertainty associated with each geolocation, and how well the confidence represents its true performance. Evaluation for this measure is quantitative and repeatable.

4.3 Satellite Image to Geospatial Data: When presented with a satellite image, whether panchromatic or multispectral, produce plausible fully attributed geospatial data representing the contents of the image. This is the equivalent of automatically generating a map or GIS data from satellite imagery and will require extensive training using fully attributed GIS data and satellite imagery. NGA suggests Open Street Map as training data for this portion, and will develop a set of globally distributed satellite images from which performance will be judged. Performance should far exceed state of the art as defined by the latest papers. As an example: https://www.mdpi.com/2220-9964/12/4/159. Because any map or set of geospatial data is an artistic human-interpretable expression of landforms found on the Earth, quantitative evaluation of this measure is nearly impossible. Some quantitative evaluation can be done by counting continuity of long or widespread features. Intersection over union can be applied to polygon features; however, at some scales, polygon features may be visualized as points and the complexity of fully attributed GIS data is so high that its utility is better judged by its analytical evaluation.

4.4 GIS Data to Geolocations: When presented with a set of spatially related points lines and polygons without a geographic reference, determine the most highly correlated match of georeferenced data. This is the equivalent of a subgraph match where the GIS data in the query information and the GIS data with which the model was trained will not match exactly. Some examples of research in this area are https://www.mdpi.com/1424-8220/23/11/5098 and https://www.frontiersin.org/articles/10.3389/feart.2023.1081445/full. The model will need to generalize the input GIS data to be able to select the area of the Earth where the GIS data is the closest match with geolocated reference data holdings. The effectiveness of the tool as measured by this metric will be how close the derived geolocations are to the ground truth, the area of uncertainty associated with each geolocation, and how well the confidence represents its true performance. NGA anticipates being able to generate a large test data set to evaluate performance. The ground truth geolocations of the GIS queries will not be released as part of the BAA. Evaluation for this measure is quantitative and repeatable.

4.5 Geospatial Data or Satellite Imagery to Ground Photos from that Location: When presented with georeferenced geospatial data or satellite imagery, generate plausible ground photos from that location. This will require training with satellite images, ground images, and worldwide geospatial data coverage so plausible imagery can be deduced based on latitude, longitude, local landforms, climate and population density. Because ground imagery is so rich and varied in its contents, NGA is unaware of any quantitative measure of the plausibility of an image. This can only be judged by comparison with other systems like Dall-E3 and earlier versions of the GFM system itself. Therefore, evaluation for this measure is qualitative judgement by the NGA team or an evaluation team under a separate contract.

4.6 GIS Data to Satellite Images in Various Modalities: When presented with georeferenced GIS data or a map, predict how satellite images may appear in various modalities, in various seasons. This will involve sufficient training with a wide variety of satellite sensors at different resolutions and seasons to convert from one modality to another. Essentially, the network envisioned will need to be of an appropriate scale to “learn” general characteristics of the Earth to such a degree that it will be able to predict how that area of the Earth would appear in the other modalities and seasons. NGA anticipates that putting together a large test data set from existing holdings will be fairly simple. Generated images can be spectrally compared against ground truth imagery from the appropriate season for evaluation. Because image generation from GIS data is also a form of digital artwork, the evaluation will be partly qualitative and performed by the NGA team or an evaluation team under a separate contract.

Deliverable Items The following table provides a minimal list of Deliverables; Appendix 1 provides the complete Contract Requirements Data List (CDRL) for each deliverable.

Table 3. Deliverables

Base Year (Proof of Concept)

Option Year 1 (Optimization)

Option Year 2 (Component Integration)

Option Year 3 (Test and Evaluation)

Intellectual Property Considerations NGA anticipates that achieving the goals of the program will necessitate a minimum of Government Purpose Rights in all deliverables. This includes all algorithms, prototype software, documentation, and training videos produced by a performer for this effort.

Notwithstanding these assumptions, the Government acknowledges the possibility that other approaches proposed may also achieve program goals. In that case, refer to section 4.5.7.3 (2), Data Rights, of the BIG-ST BAA General Solicitation for additional instructions.

Government Furnished Information/Data/Property NGA will provide performers with data required to support addressing this challenge. Additionally, the performers are invited to supply any of their own data that they have already collected from open sources, or created internally, that they may find useful. Open sources must be used at the risk of the recipient and deliverables cannot incorporate third-party-copyrighted materials without permission of the CO or the copyright owner.

The Government intends to make available:

· Maxar Vivid data as downloadable tiles via an internal contract: https://www.maxar.com/press-releases/maxar-s-high-resolution-vivid-basemaps-enhances-esri-arcgis-living-atlas-of-the-world

· TanDEM-X terrain data https://earth.esa.int/eogateway/missions/terrasar-x-and-tandem-x

· Landscan data https://landscan.ornl.gov/

Table 4. Government Furnished Information/Data/Property

Maxar VIVID or equivalent)

Security Considerations Work under this effort will be up to the Controlled Unclassified Information (CUI) level. Therefore, the minimum required security consideration for performer personnel is HSPD-12 adjudication, or higher clearance, at the beginning of the Period of Performance.

Note: Any work on classified networks (for purposes of test and integration) will be conducted by NGA.

1.1 Unclassified Work Performance Security Requirements

Uncleared Contractor personnel are authorized to work on this contract at the Unclassified level, with access up to DoD CUI level at the Contractor site without the requirement of a security clearance.

1.2 Uncleared Personnel Security Requirements

Any Contractor personnel working with CUI information must receive a favorable HSPD-12 and/or HSPD-12 Tier 1 adjudication prior to accessing CUI information. Contractor personnel who require access to CUI for 60 days or less must receive a favorable HSPD-12 adjudication. Contractor personnel who require CUI access for more than 60 days must receive a favorable HSPD-12 Tier 1 adjudication. NGA will sponsor the HSPD-12 and HSPD-12 Tier 1 background investigation for Contractor personnel, as needed.

Foreign nationals are not permitted to perform unclassified work under the terms of this contract without prior written approval from the Contracting Officer (CO) or the Contracting Officer Representative (COR). Should NGA identify the use of unauthorized personnel, the CO may direct the Contractor, at its own expense, to remove and replace any unauthorized Contractors, Subcontractors, or other personnel performing on the contract. Such action may be taken at NGA's discretion, without prejudice, to its rights under any other contract provision, e.g., termination for default.

Uncleared Contractor personnel visiting NGA facilities or other sites may receive the appropriate visitor badge and be escorted, as appropriate. The Contractor will return the visitor badge at the end of each visit day.

1.3 Physical Security Requirements

Contractor personnel are forbidden from bringing prohibited or unauthorized items into any NGA installation or other secure facility covered under this contract. These items include weapons, cell phones, cameras, two-way pagers, laptops, recording devices of any kind, flash drives, or any other removable media. Exceptions may be granted by NGA Security upon the Contractor’s request. If granted, Contractor personnel must bring documentation showing approval prior to entering with the item(s). Violations may subject the Contractor and its offending personnel to civil and/or criminal liability IAW applicable laws and regulations governing access to secure Government facilities.

1.4 Information Security Requirements

NGA shall have the sole authority to determine whether, and to what extent, protected information shall be provided to the Contractor under the terms of the contract. The Contractor shall not access, download, print, or further disseminate any unclassified information outside the execution of the defined contract requirements without the guidance and written permission from the COR.

When the Contractor receives protected information under the terms of this contract, the Contractor will comply with all applicable NGA, DoD, and IC information security policies for the proper marking, handling, processing, storing, and safeguarding of unclassified and CUI material. The Contractor will ensure that document markings are given the lowest possible security classification to maximize dissemination while still maintaining the information’s confidentiality and integrity, as necessary.

Access to CUI will be pursuant to Defense Federal Acquisition Regulation Supplement (DFARS) clause 252.204.7012. The prime contractor shall include this clause, including paragraph (m), in subcontracts, or similar contractual instruments, for operationally critical support or for which subcontract performance will involve covered defense information, including subcontracts for commercial items, without alteration, except to identify the parties. The Contractor shall determine if the information required for Subcontractor performance retains its identity as covered defense information and will require protection under this clause. Contractor personnel shall not release any unclassified information, regardless of medium (e.g., film, tape, document), pertaining to any part of this contract or any program related to this contract, unless the COR has given prior written approval or in performance of a project that has been scoped and negotiated by NGA.

NGA will assess and authorize Contractors to process unclassified information on government-owned information systems, conduct oversight of such information system processing, and provide information system security guidelines IAW with FAR clause 52.204-21 and other information system security control policies, standards, and procedures. Contractor will be responsible for executing specific provisions under the contract for the accreditation of classified information systems and, upon request, shall provide the COR a written statement confirming accreditation compliance with the minimum information systems requirements established in the DD254. The Contractor shall include the substance of FAR clause 52.204-21 in subcontracts under this contract (including subcontracts for the acquisition of commercial products or commercial services, other than commercially available off-the-shelf items), in which the Subcontractor may have federal contract information residing in or transiting through its information system.

In accordance with FAR Part 4.7 and FAR clause 52.215-2, the Contractor may retain one archival copy of any Government-furnished information provided to the Contractor during the course of performance of the contract, subject to the Contractor following relevant data safeguarding practices and CUI restrictions. The disposition of such Government-furnished information is otherwise solely at the discretion of the Government. As directed by the NGA CO or COR, the Contractor will destroy or return to NGA all Government-furnished information upon contract termination or when no longer required for contract performance.

1.5 Cybersecurity Requirements

The Contractor must, at a minimum, implement 10 U.S.C. Sections 391 and 393 along with National Institute of Standards and Technology (NIST) Special Publication (SP) 800-171, “Protecting Controlled Unclassified Information in Nonfederal Information Systems and Organizations” requirements.

If the Contractor intends to use an external cloud service provider to store, process, or transmit any covered defense information in performance of this contract, the Contractor shall require and ensure that the cloud service provider meets security requirements equivalent to those established by the Government for the Federal Risk and Authorization Management Program (FedRAMP) Moderate baseline (https://www.fedramp.gov/resources/documents/) In the event that the Contractor discovers a cyber incident that affects a covered Contractor information system or covered defense information residing therein, or that affects the Contractor’s ability to perform the requirements of the contract that are designated as operationally critical support and identified in the contract, the Contractor shall:

In the event the Contractor discovers and isolates malicious software in connection with a reported cyber incident, the Contractor shall submit the malicious software to DC3 in accordance with instructions provided by the COR. The Contractor should NOT send the malicious software to the COR.

Proposal-Specific Protection Plan This Topic Call requires a Proposal-Specific Protection Plan (PSPP), which shall be delivered with the proposal. PSPP submission information can be found in the BIG-ST BAA General Solicitation Section 4.5.8.1 and Attachment 7 of Appendix 2, BIG-ST BAA Proposal Submission Template. The PSPP shall specifically describe the protections afforded to the untrained multidimensional neural network architecture and the fully-trained foundational model.

Performance

10.1 . Place of Performance:

Performance shall take place in a performer facility anywhere in the United States.

10.2. Period of Performance (PoP):

Base Period (12 months) Option Period 1 (12 months) Option Period 2 (12 months) Option Period 3 (12 months)

Clauses All provisions and clauses from ‘Provisions_Clauses_BIGST BAA_HM047623BAA0001’ attached to the BIG-ST BAA General Solicitation are applicable to this Topic Call. The following provision is also applicable to this Topic Call:

5X52.209.9004 Use of Contractor Support

(a) The Government will use Culmen International, for administrative and/or technical support during source selection and during the term of any resulting contract. The exclusive responsibility for source selection will reside with the Government.

(b) Proprietary information submitted in response to this solicitation will be protected from unauthorized disclosure as required by Subsection 27 of the Office of Procurement Policy Act as amended (the Procurement Integrity Act) (41 U.S.C. 423) (hereinafter referred to as "the Act") as implemented in the FAR. Culmen International are bound contractually by the Protection of Information and Nondisclosure Agreements clause with respect to proprietary information. By submitting a proposal in response to this solicitation, the Contractor consents to the Government disclosing Contractor proprietary information to the aforementioned contractors for the purposes indicated.

(End of Provision)

Appendix 1 - Contract Requirements Data List (CDRL)

UNCLASSIFIED

UNCLASSIFIED

image1.jpeg image2.jpeg image3.jpeg image4.jpeg image5.jpeg image6.jpeg image7.jpeg image8.jpeg image9.jpeg image10.jpeg image11.emf

BIGST_BAA_T1_CDRL_A001_Kick-off.docx

CONTRACT DATA REQUIREMENTS LIST

(1 Data Item)

Form Approved

OMB No. 0704-0188

The public reporting burden for this collection of information is estimated to average 110 hours per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing the burden, to the Department of Defense, Executive Services Directorate

(0704-0188). Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. Please do not return your form to the above organization. Send completed form to the Government Issuing Contracting Officer for the Contract/PR No. listed in Block E.

A. CONTRACT LINE ITEM NO.

X001

B. EXHIBIT

A

C. CATEGORY:

TDP TM OTHER X-GINT

D. SYSTEM/ITEM

BIG-ST BAA

E. CONTRACT/PR NO.

HM0476-23-BAA-0001

F. CONTRACTOR

1. DATA ITEM NO.

A001

2. TITLE OF DATA ITEM

Kick-off Meeting

3. SUBTITLE

4. AUTHORITY (Data Acquisition Document No.)

Contractor Proposed, Govt Approved

5. CONTRACT REFERENCE

Sow Section XXXXX

6. REQUIRING OFFICE

7. DD 250 REQ

No

9. DIST STATEMENT REQUIRED

D see Block 16

10. FREQUENCY

ONE/R

12. DATE OF FIRST SUBMISSION

Block 16

14. DISTRIBUTION

a. ADDRESSEE

b. COPIES

8. APP CODE

N/A

11. AS OF DATE

13. DATE OF SUBSEQUENT SUBMISSION

Draft

Final

Reg

Repro

16. REMARKS

Block 9: Administrative or Operational Use (MM/DD/YYYY) Block 11: Meeting to be held NLT 10 days after start of Base Year.

Block 12: NLT 10 days after start of Base Year Period of Performance (POP).

Block 14: Deliver agenda through email, read-ahead materials, or DOD SAFE, in MS Office format at least one day prior to the scheduled meeting.

The Kick-off meeting agenda items will include, as a minimum, the following agenda items:

1) Project/contract overview including scope, objectives, deliverables and schedule, assumptions, stakeholder meetings and information requirements.

2) Potential issues, risks, questions, and concerns.

3) Government Furnished Information (GFI); and Equipment (GFE).

4) PM Office and Proposer organization, and Points of Contact information.

5) Staffing matrix identifying roles of personnel (Identify team members and roles).

6) Mutual system that the PM and Proposer will use to manage Project/Contract.

7) Required documentation; reporting requirements.

8) Funding and invoicing procedures.

9) Other agenda items, as agreed and determined by the PM Office and the Proposer.

PM

COR

15. TOTAL

G. PREPARED BY

H. DATE

I. APPROVED BY

J. DATE

18. ESTIMATED TOTAL PRICE

17. PRICE GROUP

UNCLASSIFIED//FOUO

SOURCE SELECTION SENSITIVE

DD FORM 1423-1, FEB 2001

PREVIOUS EDITION MAY BE USED.

Page of Pages

Adobe Professional 8.0

CONTRACT DATA REQUIREMENTS LIST

(1 Data Item)

A. CONTRACT LINE ITEM NO.

X001

B. EXHIBIT

A

C. CATEGORY: X-GINT

TDP TM OTHER

D. SYSTEM/ITEM

BIG-ST BAA

E. CONTRACT/PR NO.

HM0476-23-BAA-0001

F. CONTRACTOR

16. REMARKS (Continued)

DD FORM 1423-1, FEB 2001Reset

Page of Pages

INSTRUCTIONS FOR COMPLETING DD FORM 1423

(See DoD 5010.12-M for detailed instructions.)

FOR GOVERNMENT PERSONNEL

Item A. Self-explanatory.

Item B. Self-explanatory.

Item C. Mark (X) appropriate category: TDP - Technical Data Package; TM - Technical Manual; Other - other category of data, such as "Provisioning,""Configuration Management," etc.

Item D. Enter name of system/item being acquired that data will support.

Item E. Self-explanatory (to be filled in after contract award). Item F. Self-explanatory (to be filled in after contract award). Item G. Signature of preparer of CDRL.

Item H. Date CDRL was prepared.

Item I. Signature of CDRL approval authority.

Item J. Date CDRL was approved.

Item 1. See DoD FAR Supplement Subpart 4.71 for proper numbering.

Item 2. Enter title as it appears on data acquisition document cited in Item 4.

Item 3. Enter subtitle of data item for further definition of data item (optional entry).

Item 4. Enter Data Item Description (DID) number, military specification number, or military standard number listed in DoD 5010.12-L (AMSDL), or one-time DID number, that defines data content and format requirements.

Item 5. Enter reference to tasking in contract that generates requirement for the data item (e.g., Statement of Work paragraph number).

Item 6. Enter technical office responsible for ensuring adequacy of the data item.

Item 7. Specify requirement for inspection/acceptance of the data item by the Government.

Item 8. Specify requirement for approval of a draft before preparation of the final data item.

Item 9. For technical data, specify requirement for contractor to mark the appropriate distribution statement on the data (ref. DoDD 5230.24).

Item 10. Specify number of times data items are to be delivered.

Item 11. Specify as-of date of data item, when applicable.

Item 12. Specify when first submittal is required.

Item 13. Specify when subsequent submittals are required, when applicable.

Item 14. Enter addressees and number of draft/final copies to be delivered to each addressee. Explain reproducible copies in Item 16.

Item 15. Enter total number of draft/final copies to be delivered.

Item 16. Use for additional/clarifying information for Items 1 through 15. Examples are: Tailoring of documents cited in Item 4; Clarification of submittal dates in Items 12 and 13; Explanation of reproducible copies in Item 14.; Desired medium for delivery of the data item.

FOR THE CONTRACTOR

Item 17. Specify appropriate price group from one of the following groups of effort in developing estimated prices for each data item listed on the DD Form 1423.

a. Group I. Definition - Data which is not otherwise essential to the contractor's performance of the primary contracted effort (production, development, testing, and administration) but which is required by DD Form 1423.

Estimated Price - Costs to be included under Group I are those applicable to preparing and assembling the data item in conformance with Government requirements, and the administration and other expenses related to reproducing and delivering such data items to the Government.

b. Group II. Definition - Data which is essential to the performance of the primary contracted effort but the contractor is required to perform additional work to conform to Government requirements with regard to depth of content, format, frequency of submittal, preparation, control, or quality of the data item.

Estimated Price - Costs to be included under Group II are those incurred over and above the cost of the essential data item without conforming to Government requirements, and the administrative and other expenses related to reproducing and delivering such data item to the Government.

c. Group III. Definition - Data which the contractor must develop for his internal use in performance of the primary contracted effort and does not require any substantial change to conform to Government requirements with regard to depth of…