masterleopold commited on
Commit
db34faa
·
verified ·
1 Parent(s): 3f3f07e

Document Tancho Reasoner research checkpoint and staged release plan

Browse files

Private documentation only: model card, source attribution, license reference, independent-event evaluation protocol and measured inference summary. No weights, media, source code or internal run archives.

EVALUATION_PLAN.md ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Independent-event evaluation protocol v1
2
+
3
+ Prepared September 7, 2026, before accessing the reserved footage or running predictions. **Status: protocol prepared; media, reference labels and evaluation results do not yet exist.** This small pilot will not establish operational safety or broad statistical generalization.
4
+
5
+ ## Frozen scope
6
+
7
+ Compare the original Cosmos3-Edge Reasoner against the existing parent-40 export. Base revision: `a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba`. Framework revision: `0fe093eafbac390ffb7906a6f4ed9f558fe42c4c`. Parent expected weight-table SHA-256: `5fdc2753a7a5a921d9371707299f94bea75b45610d67cc3029f0d8e235c89726`. No additional training, candidate selection or prompt tuning on this event.
8
+
9
+ Use `observation-json-en-v2`, `edge-non-thinking-empty-prefix-v1`, seed 42, greedy decoding, 768 maximum new tokens and the shared canonical-PNG input representation. Verify identical inputs, preprocessing tensors and generation settings between models. A proposed baseline path must pass the same canonical-input checks as parent-40 before results count; base canonical-PNG parity is not yet an executed result.
10
+
11
+ ## Event and sampling, fixed before predictions
12
+
13
+ The entire `east-japan-typhoon-2019` event remains test-only, including related GSI flights around the Chikuma River. Initial pilot source: GSI's Nagano City Omachi UAV footage dated October 13, 2019, linked from [the official disaster page](https://www.gsi.go.jp/BOUSAI/R1.taihuu19gou.html). The reserved media URL is `https://www.gsi.go.jp/common/000219420.mp4`. Other flights are reserved, not silently substituted or used for training.
14
+
15
+ After source-specific terms review, retrieve that source, record its full SHA-256, duration and decode status, then select 12 eight-second windows mechanically. For duration D >= 96 seconds, start window i (i=0..11) at `(D - 8) * i / 11`. Use frame offsets 0.5, 2.5, 4.5 and 6.5 seconds. Record exact selected timestamps and PNG hashes. This selection must not be altered after inspecting model output. If duration is shorter, decoding fails or windows overlap, stop and version an amended sampling plan before any predictions.
16
+
17
+ Keep poor visibility, title cards and unanswerable samples in the denominator; label the input's limitations. Do not replace hard examples or force balanced decision classes. Twelve clips from one flight are correlated samples from one event, not twelve independent disaster sites. This source is a deliberately chosen pilot, not a probability sample.
18
+
19
+ ## Reference creation before inference
20
+
21
+ For each sampled window, prepare two neutral observation goals: one about visible surface/structure evidence and one about information that the images cannot support. Final goals and acceptable answers must be written and human-reviewed before either model's predictions are exposed. Avoid goal text that states the answer, source title, damage diagnosis or expected decision.
22
+
23
+ Record findings against F1–F4, uncertainty reasons and the purpose of any requested additional observation. Do not invent metric scale, unseen structural integrity, camera pose or flight permissions. Use `no_action` only for satisfaction of the stated observation goal. Freeze the 24 goal/reference records, media, prompt, source grouping and annotation-review receipt together by hash. The annotator's review is a new task; earlier approval of the 40 controls and 12 training contrasts does not cover these new references. If no independent reviewer is available, disclose developer review rather than claiming blind expert validation.
24
+
25
+ Only then run each model once on each of the 24 goals under real-image, missing-image and uniform-gray conditions: 72 planned outputs per model, 144 total. Predictions must never be used to revise the frozen reference answers. Keep audit-backed technical failures visible; any rerun is separately documented, not a replacement chosen for a better answer.
26
+
27
+ ## Metrics and review
28
+
29
+ Report separately, with counts and denominators:
30
+
31
+ 1. Raw-model JSON validity, missing/error/timeout counts and decision confusion matrix; invalid output stays in the denominator.
32
+ 2. Per-class recall, macro recall with class support, and always-inspect / always-abstain decision baselines. Report absent classes as unavailable, not zero or inferred success.
33
+ 3. Finding-level grounding: every claimed fact checked against its referenced frames; uncertainty reasons and next-observation usefulness assessed separately. Present paired model outputs with model identity hidden and ordering randomized for review.
34
+ 4. Missing-image and gray-image empty-findings-plus-abstention rates. Application-side rejection rates are a separate systems result and cannot be substituted for raw-model scores.
35
+ 5. Startup, preprocessing, native generation, output tokens and HTTP timing separately. In-container timing is not Japan network or Jetson latency. A small quality pilot is not a service-performance benchmark.
36
+
37
+ The existing development criteria (100% JSON, >=75% real-image decision agreement, macro recall >=2/3, each supported class >=50%, nonempty-grounded-observation retention and complete missing/gray abstention) remain development screens, not certified deployment thresholds. Passing mechanical screens cannot substitute for semantic review. Failure is publishable research evidence, but must not be marketed as an operationally validated model.
38
+
39
+ ## Release and execution dependencies
40
+
41
+ Before running paid inference, fix a concrete GPU job, current price, time limits, result recovery and GPU/disk deletion path. This protocol does not queue a job or authorize an unspecified budget. Before opening the reserved media in data-building code, create a versioned evaluation-only event policy and acquisition manifest; preserve the current unopened policy as historical evidence and prohibit train/validation access.
42
+
43
+ Publish the protocol, actual deviations, counts, failures and review status with results. Once the event has informed model or prompt changes, it becomes development evidence for those successors; reserve a different event for their next independent evaluation. A public research release additionally depends on rights/notice review and an independently reproducible loading package.
LICENSE-DOCUMENTATION ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Tancho contributors
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
NOTICE.md ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Licensing and provenance notice
2
+
3
+ This repository currently contains original Tancho documentation and a reference copy of the upstream model license. It does not distribute NVIDIA model weights or source code, training videos, extracted frames, labels, credentials or internal run archives.
4
+
5
+ ## Original documentation
6
+
7
+ The original Tancho documentation in this preparation repository is covered by [LICENSE-DOCUMENTATION](LICENSE-DOCUMENTATION), the project's MIT license. This documentation license grants no rights in model weights, third-party media, trademarks or separately licensed material. It is not an assignment of joint-project intellectual property.
8
+
9
+ ## NVIDIA model and software
10
+
11
+ The base is [NVIDIA Cosmos3-Edge](https://huggingface.co/nvidia/Cosmos3-Edge), revision `a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba`. Its model card identifies [OpenMDW 1.1](https://openmdw.ai/license/). A reference copy is at [licenses/LICENSE.OpenMDW-1.1](licenses/LICENSE.OpenMDW-1.1). A future redistribution must preserve that license and all applicable copyright and origin notices from the artifacts actually distributed. This reference copy alone does not constitute a complete weight-package clearance.
12
+
13
+ The training runtime used NVIDIA Cosmos Framework revision `0fe093eafbac390ffb7906a6f4ed9f558fe42c4c`. Any future software/container distribution must inventory and retain its own NVIDIA and dependency notices. The Tancho code repository's MIT license does not replace upstream terms. No NVIDIA or Nebius endorsement is asserted.
14
+
15
+ ## Media and annotations
16
+
17
+ The source footage is attributed to the Geospatial Information Authority of Japan (GSI). Refer to [SOURCES.md](SOURCES.md). GSI applies PDL 1.0 unless otherwise stated and requires attribution, identification of editing/processing, and checks of third-party rights and individual conditions. Public availability alone does not establish blanket clearance for redistribution.
18
+
19
+ Tancho's clipping, resizing, observation goals and annotations are project-created derivatives, not official GSI labels or assessments. Human approval of label correctness is distinct from approval to redistribute media or joint-project deliverables. No media or annotations are included in this repository.
20
+
21
+ ## Publication authority
22
+
23
+ The user selected Vox's `v13s` Hugging Face namespace for this private preparation repository. That selection establishes the destination; it does not resolve ownership, contractual approval or use of Taisei's name for a subsequent public release. Public model/data distribution remains pending that determination.
README.md ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ - ja
5
+ base_model: nvidia/Cosmos3-Edge
6
+ base_model_relation: finetune
7
+ tags:
8
+ - physical-ai
9
+ - disaster-observation
10
+ - research-preview
11
+ - cosmos3
12
+ - nebius
13
+ ---
14
+
15
+ # Tancho Reasoner — research preview
16
+
17
+ **Status: private preparation repository; documentation only. No model weights, training footage or executable model package are uploaded here.** Updated September 7, 2026.
18
+
19
+ Tancho is a research prototype for describing visible evidence and missing information in aerial disaster observations. It fine-tunes the Reasoner component of NVIDIA Cosmos3-Edge on Nebius AI Cloud. The intended publishing namespace is [Vox (`v13s`)](https://huggingface.co/v13s). This page does not establish ownership of joint-project deliverables or imply NVIDIA, Nebius or GSI endorsement.
20
+
21
+ The trained research checkpoint exists and has been reloaded and used for GPU inference. Its generalization, factual grounding and suitability for operational disaster response are not established. It does not provide flight commands or a validated camera-trajectory policy.
22
+
23
+ ## Model and training
24
+
25
+ | Item | Recorded configuration |
26
+ |---|---|
27
+ | Base | `nvidia/Cosmos3-Edge` |
28
+ | Base revision | `a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba` |
29
+ | NVIDIA Cosmos Framework revision | `0fe093eafbac390ffb7906a6f4ed9f558fe42c4c` |
30
+ | Candidate described | Parent-40 Reasoner checkpoint, trained September 6, 2026 |
31
+ | Method | Supervised fine-tuning of language/connecting parameters; frozen visual encoder; not LoRA |
32
+ | Trainable parameters | 2,022,966,784 across 233 tensors |
33
+ | Frozen visual parameters | 412,649,712 across 437 tensors |
34
+ | Data actually consumed | 20 reviewed Aso Bridge observation examples, two passes, 40 updates |
35
+ | Training input | Four frames from each short clip, 0.5 fps; English task prompt and Japanese target JSON |
36
+ | Optimizer | Native FusedAdam in AdamW mode, FP32 parameters, BF16 forward, SDPA, microbatch 1 |
37
+ | Learning rate | Maximum 1e-6; two warmup steps; 40-step schedule |
38
+ | Hardware | One NVIDIA L40S on Nebius AI Cloud |
39
+ | Saved artifacts | Native distributed checkpoints and BF16 HF Reasoner export, 670 tensors, approximately 4.9 GB |
40
+ | Expected weight-table SHA-256 | `5fdc2753a7a5a921d9371707299f94bea75b45610d67cc3029f0d8e235c89726` |
41
+
42
+ The weight-table digest identifies the internal expected tensor-hash table; it is not a downloadable checkpoint-file checksum or a cryptographic signature. Later 48- and 80-update candidates were rejected after development diagnostics. The 44-example canonical-PNG pack was used in later preprocessing checks, not in the parent-40 training run. VideoPhy was earlier infrastructure preparation, not the parent-40 training corpus.
43
+
44
+ ## Input, output and intended use
45
+
46
+ Input is an observation goal and four timestamp-ordered images, referenced as F1–F4. Output is Japanese JSON with `findings` (text and evidence references), `unknowns`, `decision`, and `next_observation`.
47
+
48
+ - `inspect`: request additional information relevant to the observation goal.
49
+ - `no_action`: the limited observation goal is already satisfied; this is not a declaration of safety or absence of damage.
50
+ - `abstain`: the requested conclusion cannot be supported from the input.
51
+
52
+ Intended use is offline research, failure analysis and demonstrations with human assessment. Structural validity and evidence IDs do not prove that an observation is true. The application separately rejects missing, unreadable or uniform images and keeps generated responses pending review. Those application checks are not improvements to the underlying model's scores.
53
+
54
+ ## Development evaluation
55
+
56
+ The September 6 comparison used 12 Sannoudani River clips, each presented with real images, no images and gray images. Base and parent-40 used the same cached NVIDIA Omni inference path, input tensors, prompt (`observation-json-en-v2`), seed 42, greedy generation and 768-token limit. Invalid and missing outputs count as failures.
57
+
58
+ | Metric | Base | Parent-40 |
59
+ |---|---:|---:|
60
+ | JSON contract valid, all conditions | 2/36 | 36/36 |
61
+ | Decision agreement, real images | 0/12 | 6/12 |
62
+ | Empty findings plus abstention, no images | Not summarized here | 0/12 |
63
+ | Empty findings plus abstention, gray images | Not summarized here | 0/12 |
64
+
65
+ The real-image labels contain six `inspect`, three `no_action` and three `abstain` cases. Always predicting `inspect` also achieves 6/12. Decision agreement includes format failures and is not a pure visual-understanding score. The model sometimes reuses facts from the goal or prompt without visual evidence. The empty-findings metric is a contract check, not an expert-adjudicated hallucination rate.
66
+
67
+ These 12 clips are repeatedly inspected development data from the same disaster event as training. They are not an independent test set. The original MP4 decoding path used in this comparison differs from the later canonical-PNG path, so these quality scores do not certify that newer input path. See [the planned independent-event evaluation](EVALUATION_PLAN.md).
68
+
69
+ ## Verified inference engineering
70
+
71
+ On September 7, two GPU HTTP requests reused one parent-40 model load. All 670 loaded tensor hashes were verified at startup. A01 and A02 reproduced the earlier cold-worker input tensors and output token sequences for the same PNG inputs.
72
+
73
+ | Measurement | Value |
74
+ |---|---:|
75
+ | Model startup | 30.438 seconds |
76
+ | First request, A01, HTTP | 33.938 seconds |
77
+ | Second request, A02, HTTP | 2.449 seconds |
78
+ | Second request, A02, native generation | 2.209 seconds |
79
+
80
+ This is two requests, once each, inside a Nebius L40S container. HTTP uses container loopback; the numbers exclude Japan-to-cloud networking, VM provisioning and container-image download. They are not a latency distribution, SLA or Jetson benchmark. Both answers still await human semantic review. The temporary service has terminated; this repository does not host a live inference endpoint.
81
+
82
+ The selected measurements and source-summary digest are recorded in [evidence-summary.json](evidence-summary.json). Internal infrastructure identifiers and raw observation content are excluded from that summary.
83
+
84
+ ## Limitations and deployment status
85
+
86
+ - Very small, deliberately selected training set from one disaster site; no representative normal-site or multi-event benchmark.
87
+ - Unreliable separation of visible facts, unseen conditions and temporal change; missing-image controls failed.
88
+ - No independently reviewed generalization result, calibrated confidence or structural/geotechnical diagnosis.
89
+ - No validated Jetson runtime, camera9d generation, autonomous flight or operational disaster-response integration.
90
+ - The frozen generator and any original upstream capabilities are not validated Tancho outputs. The upstream Cosmos3-Edge model size must not be used to describe this Reasoner-only export as an entirely fine-tuned model.
91
+
92
+ Native inference currently requires the pinned Linux GPU framework and a verified Reasoner overlay. No drop-in `transformers` loading example is offered for this repository because its weights and runnable distribution are not present. A future release needs an independently reproduced installation procedure.
93
+
94
+ ## Data, licenses and release status
95
+
96
+ Parent-40 used reviewed observations derived from GSI's public 2016 Kumamoto UAV footage around Aso Bridge. Sannoudani was development evaluation; 2017 Kyushu footage was used only for later rejected candidates and development work. The reserved 2019 typhoon event has not yet been downloaded, viewed, labeled or evaluated by this project.
97
+
98
+ See [SOURCES.md](SOURCES.md) for attribution and [NOTICE.md](NOTICE.md) for distinct documentation, model and data terms. Original documentation is MIT licensed. NVIDIA Cosmos3-Edge is provided under OpenMDW 1.1; this does not make third-party footage MIT licensed or settle downstream/joint-project rights. The final model-distribution license and publication authority remain to be finalized; the model-card license metadata is intentionally unset.
99
+
100
+ The current distribution is private documentation. Proposed later stages are approved public documentation/code, followed by a research-weight release after rights and evaluation review. Project-specific operational data and customer models are outside that proposed research release.
SOURCES.md ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Data sources and scope
2
+
3
+ Source and terms review date: September 7, 2026. This file lists provenance; it redistributes no source footage or derived frames. Hashes below are computed from locally downloaded media and are not publisher-issued checksums.
4
+
5
+ | Source | Use in this project | Original media SHA-256 |
6
+ |---|---|---|
7
+ | GSI, Aso Bridge / Tateno landslide, April 16, 2016 | Parent-40 training: 20 reviewed examples, consumed twice | `3fddc185f274d0b6179879cdf36c41aafff1b2544df436ab56cf2ecab10d134c` |
8
+ | GSI, Sannoudani River landslide, April 17, 2016 | 12 repeatedly seen development-validation examples; never independent test | `4bc19fe643470ba0ae627c10618b6dca6e4ae5a6d80f79a70a812109a87335a4` |
9
+ | GSI, Kitagawa, Asakura, July 7, 2017 | Later development data and rejected follow-up candidates; not parent-40 training | `c92103739ffd2d8f025dbc33f92ec3278aa7c999505b66f33b457172793b2f73` |
10
+ | GSI, Kagetsu River bridge collapse, Hita, July 7, 2017 | Later development data and rejected follow-up candidates; not parent-40 training | `9d82df413cca0ef8fc860cec561725cff4686d3feee2d843ea0f5dd5786f4fa6` |
11
+
12
+ Official source pages:
13
+
14
+ - [2016 Kumamoto earthquake UAV observations](https://www.gsi.go.jp/BOUSAI/H27-kumamoto-earthquake-index.html).
15
+ - [2017 northern Kyushu heavy-rain UAV observations](https://www.gsi.go.jp/BOUSAI/H29hukuoka_ooita-heavyrain.html).
16
+ - [2019 East Japan typhoon observations](https://www.gsi.go.jp/BOUSAI/R1.taihuu19gou.html): reserved evaluation event, metadata only; no model evaluation has been performed.
17
+ - [GSI content terms](https://www.gsi.go.jp/kikakuchousei/kikakuchousei40182.html).
18
+
19
+ Attribution for any subsequently cleared derived examples:
20
+
21
+ > 出典:国土地理院ウェブサイト(該当する災害ページのURL)。元動画の切り出し・縮小・観測目的・注釈はTancho開発側で作成。国土地理院の公式評価ではありません。
22
+
23
+ The current [GSI terms](https://www.gsi.go.jp/kikakuchousei/kikakuchousei40182.html) apply PDL 1.0 absent a special notice. Individual-source notices, third-party rights and any applicable separate conditions still require review for the intended distribution. The source-page snapshots and media hashes in private project provenance support traceability; they are not legal approval.
24
+
25
+ Earlier VideoPhy infrastructure examples did not train the parent-40 checkpoint. No Taisei customer footage is identified in this checkpoint's recorded training corpus. This does not establish rights in the broader joint project.
evidence-summary.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": "tancho-research-evidence-1.0",
3
+ "as_of": "2026-09-07",
4
+ "checkpoint": "parent-40",
5
+ "independent_test_completed": false,
6
+ "production_validated": false,
7
+ "base_model": "nvidia/Cosmos3-Edge",
8
+ "base_revision": "a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba",
9
+ "expected_weight_table_sha256": "5fdc2753a7a5a921d9371707299f94bea75b45610d67cc3029f0d8e235c89726",
10
+ "warm_gpu": {
11
+ "hardware": "NVIDIA L40S",
12
+ "location": "Nebius AI Cloud, in-container loopback HTTP",
13
+ "startup_seconds": 30.437605734999977,
14
+ "same_worker_verified": true,
15
+ "native_preflight_samples": 44,
16
+ "preflight_was_training": false,
17
+ "semantic_truth_verified": false,
18
+ "human_reviewed": 0,
19
+ "source_summary_sha256": "b8640aa23a39aad319b28a1986f956f5a2330bf8aa716d550cb1c9c6172c91d6",
20
+ "records": [
21
+ {
22
+ "id": "A01",
23
+ "http_status": 200,
24
+ "http_seconds": 33.938056,
25
+ "generation_seconds": 3.0592310909999583,
26
+ "input_tokens": 1173,
27
+ "output_tokens": 90,
28
+ "status": "pending_human_review",
29
+ "weights_verified": true
30
+ },
31
+ {
32
+ "id": "A02",
33
+ "http_status": 200,
34
+ "http_seconds": 2.448578,
35
+ "generation_seconds": 2.2090070510000146,
36
+ "input_tokens": 1179,
37
+ "output_tokens": 86,
38
+ "status": "pending_human_review",
39
+ "weights_verified": true
40
+ }
41
+ ]
42
+ },
43
+ "scope": "Two requests once each; not a network, Jetson, production latency or semantic accuracy benchmark. The earlier quality evaluation used a different MP4 preprocessing path."
44
+ }
licenses/LICENSE.OpenMDW-1.1 ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)
2
+
3
+ By exercising rights granted to you under this agreement, you accept and agree
4
+ to its terms.
5
+
6
+ As used in this agreement, "Model Materials" means the materials provided to
7
+ you under this agreement, consisting of: (1) one or more machine learning
8
+ models (including architecture and parameters); and (2) all related artifacts
9
+ (including associated data, documentation and software) that are provided to
10
+ you hereunder.
11
+
12
+ Subject to your compliance with this agreement, permission is hereby granted,
13
+ free of charge, to deal in the Model Materials without restriction, including
14
+ under all copyright, patent, database, and trade secret rights included or
15
+ embodied therein.
16
+
17
+ If you distribute any portion of the Model Materials, you shall retain in your
18
+ distribution (1) a copy of this agreement, and (2) all copyright notices and
19
+ other notices of origin included in the Model Materials that are applicable to
20
+ your distribution.
21
+
22
+ If you file, maintain, or voluntarily participate in a lawsuit against any
23
+ person or entity asserting that the Model Materials directly or indirectly
24
+ infringe any patent or copyright, then all rights and grants made to you
25
+ hereunder are terminated, unless that lawsuit was in response to a
26
+ corresponding lawsuit first brought against you.
27
+
28
+ This agreement does not impose any restrictions or obligations with respect to
29
+ any use, modification, or sharing of any outputs generated by using the Model
30
+ Materials.
31
+
32
+ THE MODEL MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
33
+ OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
34
+ FITNESS FOR A PARTICULAR PURPOSE, TITLE, NONINFRINGEMENT, ACCURACY, OR THE
35
+ ABSENCE OF LATENT OR OTHER DEFECTS OR ERRORS, WHETHER OR NOT DISCOVERABLE, ALL
36
+ TO THE GREATEST EXTENT PERMISSIBLE UNDER APPLICABLE LAW.
37
+
38
+ YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT MAY
39
+ APPLY TO THE MODEL MATERIALS OR ANY USE THEREOF, INCLUDING WITHOUT LIMITATION
40
+ ANY PERSON'S COPYRIGHTS OR OTHER RIGHTS INCLUDED OR EMBODIED IN THE MODEL
41
+ MATERIALS; (2) OBTAINING ANY NECESSARY CONSENTS, PERMISSIONS OR OTHER RIGHTS
42
+ REQUIRED FOR ANY USE OF THE MODEL MATERIALS; OR (3) PERFORMING ANY DUE
43
+ DILIGENCE OR UNDERTAKING ANY OTHER INVESTIGATIONS INTO THE MODEL MATERIALS OR
44
+ ANYTHING INCORPORATED OR EMBODIED THEREIN.
45
+
46
+ IN NO EVENT SHALL THE PROVIDERS OF THE MODEL MATERIALS BE LIABLE FOR ANY CLAIM,
47
+ DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
48
+ OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL MATERIALS, THE
49
+ USE THEREOF OR OTHER DEALINGS THEREIN.