dfkev commited on
Commit
1ad49bc
·
verified ·
1 Parent(s): 6baa202

rewrite org card prose: drop AI-tells, replace 'security veterans'

Browse files
Files changed (1) hide show
  1. README.md +41 -39
README.md CHANGED
@@ -9,57 +9,59 @@ pinned: false
9
 
10
  # Deep Fathom
11
 
12
- **Making continuous CMMC readiness the norm for the Defense Industrial
13
- Base.**
14
 
15
- We're builders, security veterans, and engineers shipping an
16
- AI-native compliance system for CMMC one that guides priorities,
17
- orchestrates work across teams and tools, automates the routine steps,
18
- verifies in flow, and keeps documentation synchronized as the work
19
- moves.
20
 
21
- This is where we publish the open building blocks of that system.
 
 
22
 
23
- ## Active project
24
 
25
- [**basalt**](https://huggingface.co/deepfathom/basalt) a CMMC /
26
- NIST SP 800-171 / DFARS practitioner Q&A LoRA adapter on top of
27
- `ibm-granite/granite-4.1-8b`. Currently private during M-Train v1
28
- validation; release is contingent on completed eval and license review
29
- of the base-model and corpus terms.
 
30
 
31
  ## How we build the models
32
 
33
- - **Curation over scale.** ~13K judge-validated practitioner pairs
34
- beat 100K of unfiltered scrapes for narrow regulatory domains.
35
- - **Provenance everywhere.** Every published artifact ties back to a
36
- reproducible training run — data version, seed, hyperparameters,
37
- judge configuration.
38
- - **Adapters, not merges.** Domain deltas ship as LoRA adapters,
39
- composable at runtime with the upstream base via PEFT / vLLM. Users
40
- get the latest base model plus our specialty layer, swappable
41
- independently.
 
 
 
 
42
 
43
  ## Principles
44
 
45
- - **More rigor. Radically less effort.** Higher compliance standards
46
- delivered with less friction.
47
- - **Frontier AI, applied.** Where useful, where the math is real,
48
- where the trade-offs are honest.
49
- - **Security and compliance, intertwined by design.** Proof and
50
- protection, not bolted on after the fact.
51
- - **One truth, shared authority.** Same view across the team, the
52
- partner, and the assessor.
53
 
54
  ## Who we work with
55
 
56
- - **Defense contractors and DIB suppliers** from primes down to the
57
- small businesses CMMC was always going to test hardest.
58
- - **MSPs and MSSPs** delivering compliance services without building
59
- the in-house regulatory expertise from scratch.
60
- - **C3PAOs and assessors** who need shared evidence, not screenshots.
61
 
62
- ## Reach us
63
 
64
- [deepfathom.ai](https://www.deepfathom.ai) · model repo discussions
65
- for technical questions.
 
9
 
10
  # Deep Fathom
11
 
12
+ Our mission is to make continuous CMMC readiness the norm for the
13
+ Defense Industrial Base.
14
 
15
+ Deep Fathom builds an AI-native compliance platform for CMMC. The
16
+ platform takes on the routine work, tracks evidence as teams generate
17
+ it, and keeps documentation aligned with the actual state of the
18
+ controls. The people behind it are engineers and compliance
19
+ practitioners who got tired of treating audits as ceremony.
20
 
21
+ This Hugging Face org is the open side of that work. We publish the
22
+ ML pieces we can release: model adapters today, datasets later when
23
+ licensing allows.
24
 
25
+ ## What's here
26
 
27
+ [basalt](https://huggingface.co/deepfathom/basalt) is the only public
28
+ artifact right now. It is a LoRA adapter for CMMC, NIST SP 800-171,
29
+ and DFARS practitioner Q&A, layered on `ibm-granite/granite-4.1-8b`.
30
+ The repo is private through M-Train v1 validation. Public release
31
+ depends on eval results and a license review of the base model and
32
+ corpus terms.
33
 
34
  ## How we build the models
35
 
36
+ We choose curation over scale. Roughly 13,000 judge-validated
37
+ practitioner pairs do better work than 100,000 unfiltered scrapes
38
+ when the domain is this narrow.
39
+
40
+ Every artifact links back to a reproducible training run. We record
41
+ the data version, seed, hyperparameters, and judge configuration so a
42
+ later release can be checked against the artifact it claims to
43
+ supersede.
44
+
45
+ We publish LoRA adapters rather than merged weights. PEFT and vLLM
46
+ load the base and the adapter together at runtime. The result is the
47
+ latest upstream base, our specialty layer composed on top of it, and
48
+ the option to swap each layer independently.
49
 
50
  ## Principles
51
 
52
+ 1. More rigor. Radically less effort.
53
+ 2. Frontier AI, applied where it carries weight.
54
+ 3. Security and compliance treated as one workflow.
55
+ 4. One truth, shared across the team, the partner, and the assessor.
 
 
 
 
56
 
57
  ## Who we work with
58
 
59
+ Defense contractors and DIB suppliers, including the small businesses
60
+ that CMMC was always going to test hardest. MSPs and MSSPs that
61
+ deliver compliance services for clients. C3PAOs and assessors who
62
+ want shared evidence instead of screenshots.
 
63
 
64
+ ## Reach
65
 
66
+ [deepfathom.ai](https://www.deepfathom.ai). Use model repo
67
+ discussions for technical questions.