MarshallDoyle/NASA-GPT-OSS

NASA OSS Model Card

Highlights

• Fine-tune of OpenAI GPT-OSS (20B) using Unsloth for optimized training.
• Trained on synthetic Q&A data derived from all available NASA standards and handbooks (excluding center-level standards).
• Data generated via chunking into 4096 tokens with 256 overlap, question + answer pairs produced per chunk.
• Provides compliance-oriented, clause-referenced outputs for NASA engineering standards.
• Extends prior work (NASA Systems Engineering Llama-3.1 8B) to dozens of NASA standards, not just one handbook.
• Best used with retrieval-augmented generation (RAG): include the relevant standard text in the prompt for highest accuracy.

Recommended Inference Settings:

• temperature = 1
• top_k = 0
• top_p = 1

Official OpenAI GPT-OSS page: https://openai.com/oss

---

Model Summary

The NASA OSS model is designed to assist with interpretation, reasoning, and compliance guidance across NASA standards and handbooks.
It is particularly effective for:

• Engineering decision support
• Standards compliance verification
• Requirements and process traceability
• Education and training of new engineers

---

Training Data

• Sources: All publicly available NASA standards and handbooks (excluding center-level standards).
• Processing pipeline:
  • Documents chunked into 4096 tokens with 256-token overlap.
  • Each chunk used to generate multiple synthetic questions relevant only to that section.
  • Answers generated strictly from the chunk (but phrased naturally without explicit chunk references).

---

Recommended Usage

• Use with RAG: Retrieval of the intended NASA standard or handbook section and injection into the prompt significantly improves performance.
• Verification: Always cross-check outputs against the official NASA source documents before applying in mission-critical work.

---

Comparison Examples

Example 1 – Required Components for Material Process Qualification (NASA-STD-6030)

Llama-3.1 8B Fine-tune (Systems Eng only):
"Nasa standards generally require process control and material specifications before flight qualification."

GPT-OSS Base:
"Material specifications and process definitions must be established for qualification."

NASA OSS:
NASA-STD-6030 – Materials Process Qualification
The standard requires that a candidate material process be fully characterized before it can be qualified for use in additive manufacturing hardware. Three distinct components must be defined and documented:

1. Feedstock Material Specification – properties of the raw AM material.
2. Part Material Specification (or equivalent) – final part-material characteristics after processing.
3. Material Processing Specification (MPS) – process parameters and QA procedures.
   These items are listed in NASA-STD-6030 §2 f (g–i) and clarified in §8 as required for a Qualified Part Process.

---

Example 2 – PCRDs (Primary Control-Relevant Data) for Metallic AM Processes

Llama-3.1 8B Fine-tune:
"Typical values include tensile strength and elongation."

GPT-OSS Base:
"PCRDs are mechanical properties but fracture toughness is sometimes excluded."

NASA OSS:
Question: Which of the following is NOT one of the four PCRDs defined for each metallic AM MPS: ultimate tensile strength, yield strength, elongation, or fracture toughness?
Answer: Fracture toughness is NOT one of the PCRDs.
NASA-STD-6030 explicitly defines: ultimate tensile strength, yield strength, elongation, and reduction of area.
Reference: NASA-STD-6030 §4.3.2

---

Example 3 – Powder Certification and Reuse Rules for Ti-6Al-4V (LPBF)

Llama-3.1 8B Fine-tune:
"Powder must be certified before use in additive manufacturing."

GPT-OSS Base:
"Certification and reuse requirements depend on program standards."

NASA OSS:
NASA-STD-6030 requires:

• Feedstock certification for composition, cleanliness, and particle morphology (§5.1).
• Documentation of lot traceability and maximum reuse cycles (§5.2).
• Validation via witness coupon testing before production (§7.3).
  These steps ensure powder consistency and qualification evidence for flight hardware.

---

Example 4 – Wire-Feed Directed Energy Deposition (DED) Qualification

NASA OSS:
For wire-feed DED processes, NASA-STD-6030 mandates documentation of:

• Material Specification – composition, purity, heat-treatment requirements (§7.2).
• Processing Parameters – machine-specific build settings and post-processing (§7.3).
• Testing & Inspection Methods – destructive and nondestructive evaluations with acceptance criteria (§7.3).
  These form the Candidate Material Process (CMP) and serve as the foundation for establishing a Qualified Part Process (QPP).

---

Example 5 – Dimensional Inspection for AM Structural Truss

NASA OSS:
NASA-STD-6030 requires GD&T compliance verification and interface checks through the Additive Manufacturing Control Plan (AMCP).

• MPS, QMP, and AMCP integration define dimensional verification (§4.1–4.3).
• Witness coupon testing and statistical sampling ensure dimensional repeatability (§7.2–7.3).
  Reference: NASA-STD-6030, §4.2; §7.2–7.3

---

Limitations

• Model outputs reflect public NASA standards only.
• May not cover internal center-level or proprietary standards.
• Best used with retrieval context – performance drops without standard text injection.

---

Ethical Considerations

• Should be treated as an assistive tool, not as a replacement for human engineering judgment.
• Outputs must be verified against authoritative NASA documentation.
• Not suitable for export-controlled, ITAR-restricted, or classified projects.

---

Citation

If you use this model, please cite as:

@misc{marshall2025nasaoss, author = {Marshall Doyle}, title = {NASA OSS: Domain-Specific Fine-Tune of GPT OSS on NASA Standards}, year = {2025}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/MarshallDoyle/NASA-OSS}} }

---