"front"	"back"	"tags"
"Name the three tiers of model openness, from least to most open."	"(1) OPEN-WEIGHTS-ONLY — weights, no training data (Llama 3.x). (2) OPEN-DATA — weights + training corpus or reproducible pipeline (MiniCPM, OLMo, Tulu, SmolLM3). (3) OPEN-RECIPE — weights + data + full training code/config, often checkpoints (OLMo, Tulu 3, SmolLM3)."	c3::ft02::recall
"What does each openness tier let you DO that the one below does not?"	"Weights-only lets you USE the model. Open-data lets you AUDIT what it saw. Open-recipe lets you REPRODUCE and PROVE the whole thing. Each tier adds a trust property the one below lacks."	c3::ft02::recall
"What is the single-sentence distinction between weights-only and open-data?"	"Open-weights-only = you trust the publisher's word about what the model saw. Open-data = you can PROVE what the model saw. That auditability difference is load-bearing for HIPAA / IL5 / IL6 / air-gapped."	c3::ft02::application
"What is OSAID v1.0, when was it released, and where?"	"The Open Source AI Definition, published by the Open Source Initiative (OSI). Version 1.0 was released October 28, 2024 (2024-10-28) at the All Things Open (ATO) conference. The first industry-standardized definition of 'open source' for AI systems."	c3::ft02::recall
"What is the OSAID 'data clause' and why is it a deliberate compromise?"	"OSAID requires 'sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.' It does NOT require the data itself, nor an open license on the data. This is a compromise so that releases trained on private/licensed/personal data (which can't legally be redistributed) can still claim openness."	c3::ft02::recall
"What is 'the OSAID gap,' and why must you not conflate it with reproducibility?"	"The OSAID gap is the deliberate distance between OSAID compliance and actual reproducibility/auditability. A Llama-style release (weights + described-in-aggregate data, no corpus) can be OSAID-compliant WITHOUT being reproducible. 'OSAID-compliant' and 'open-data/recipe' are different claims — never conflate them."	c3::ft02::analysis
"What does the NTIA 2024 Open-Model Weights Report say about sensitive data, and why is it the key citation?"	"It states open-weight models 'provide security benefits by allowing firms, researchers, and users to use potentially sensitive data' locally / on-premises. It's the single best GOVERNMENT citation for the sensitive-data/on-prem argument. Federal agency, official report mandated under EO 14110, July 2024."	c3::ft02::recall
"Why does the NTIA argument make open-weights a prerequisite for IL5/IL6 and air-gapped deployments?"	"If weights aren't available, the model can only run via a vendor API — every inference leaves your trust boundary. You can't put a proprietary API model inside a SCIF/air-gap. Open weights let the model run inside the boundary; open-data/recipe additionally lets you answer 'what did it see?'. (Sets up FT22.)"	c3::ft02::analysis
"What is the Stanford FMTI and what were its May 2024 headline numbers?"	"The Foundation Model Transparency Index (Stanford CRFM; Bommasani et al., arXiv:2407.12929). Scores developers on 100 transparency indicators. May 2024: mean score 58/100, top score 85/100. Fully-open developers cluster at the top; closed developers (OpenAI ~49, Anthropic ~51) below the mean."	c3::ft02::recall
"How does the FMTI relate to the openness tiers?"	"It tracks them cleanly: open-recipe developers (OLMo, IBM Granite) score highest; weights-only lower; closed lowest. The FMTI turns 'openness is nice' into a measurable, citable procurement criterion — the bridge from slogan to number."	c3::ft02::analysis
"Classify these on the open spectrum: MiniCPM, OLMo 2, Tulu 3, SmolLM3, Nemotron, DCLM, Llama 3.1, GPT-4o."	"OPEN-RECIPE: MiniCPM, OLMo 2, Tulu 3, SmolLM3, DCLM. OPEN-WEIGHTS/PARTIAL: Nemotron (post-train data documented). OPEN-WEIGHTS-ONLY: Llama 3.1 (data described in aggregate only). CLOSED: GPT-4o (weights withheld, API-only)."	c3::ft02::application
"Which model families are OPEN-DATA or OPEN-RECIPE (the auditable ones)?"	"OpenBMB MiniCPM (Apache-2.0, Ultra* datasets), Allen Institute OLMo 2 and Tulu 3 (Apache-2.0, fully open), HuggingFace SmolLM3 (Apache-2.0, full recipe), DCLM (data pipeline released). All auditable."	c3::ft02::recall
"Which releases are OPEN-WEIGHTS-ONLY or CLOSED (not auditable)?"	"Meta Llama 3.x incl. 405B — weights under the Llama Community License, data described in aggregate, recipe summarized. OpenAI GPT-4o — closed, weights withheld, API-only, proprietary. Neither can answer 'what did the model see?' from primary sources."	c3::ft02::recall
"Who is OpenBMB, and what is the MiniCPM family?"	"OpenBMB = 'Open Lab for Big Model Base,' a collaboration between Tsinghua University and the company ModelBest. The MiniCPM family includes MiniCPM5-1B, MiniCPM3-4B, MiniCPM-V 4.6 (vision), and MiniCPM-o 4.5 (omni). Apache-2.0. The course's on-ramp hero — runs/fine-tunes on consumer hardware."	c3::ft02::recall
"What open datasets does OpenBMB ship alongside MiniCPM, and why do they matter to this course?"	"UltraChat (large-scale dialogue), UltraFeedback (preference/feedback data), and Ultra-FineWeb (curated web pretraining mix). They are open, documented, and reusable — the default examples the data modules (Pillar 1) return to for open preference/pretraining data."	c3::ft02::application
"Why is MiniCPM5-1B the 'on-ramp hero' of the course?"	"It sits in a sweet spot: genuinely open-data/open-recipe under Apache-2.0, small enough to run and fine-tune on a single consumer GPU, and accompanied by reusable open datasets. It's the auditable base the early modules load — you can point at every byte it saw."	c3::ft02::analysis
"State the three properties that only open-data/open-recipe releases give you for sensitive domains, and the compliance requirement each maps to."	"(1) AUDITABILITY — prove what the model saw (HIPAA risk analysis). (2) REPRODUCIBILITY — rebuild years later, prove no silent drift (validated-pipeline integrity). (3) SUPPLY-CHAIN TRUST — rule out hidden training-time exfiltration (IL5/IL6, air-gap). Each maps to a requirement a regulator can name."	c3::ft02::analysis
"What is 'silent drift,' and how does open-recipe prevent it?"	"Silent drift = a vendor silently updates a closed model, changing behavior underneath your validated pipeline so your validation no longer holds. Open-recipe lets you pin the exact commit of data+code+weights and rebuild on demand — you can prove the model you validated is the model you're running."	c3::ft02::analysis
"Why is 'it's open, so it's safe' an anti-pattern?"	"Open data is AUDITABLE, not SAFE. An open corpus can still contain PII, copyrighted text, poisoned examples, or liability-creating material. Openness gives you the ABILITY to vet; it does not do the vetting. Open bases still need a data audit (FT04-07), PII sweep, licensing review, and red-teaming."	c3::ft02::analysis
"Why is 'trusting a community merge with no provenance' an anti-pattern?"	"HuggingFace Hub merges often have no documented data lineage, no evals, and a license the uploader probably can't grant. A model card saying 'merge of X and Y, works great' is not an audit. In sensitive deployment, trace provenance to the original bases and data — if you can't, treat it as unauditable."	c3::ft02::application
"What does the OLMo 2 paper (arXiv:2501.00656) mean by 'fully open,' and why is it the canonical citation?"	"Fully open = data + code + weights + intermediate checkpoints + evaluation suite, all released. OLMo 2 (Allen Institute for AI) is the canonical open-recipe citation because it ships every component a skilled person needs to rebuild the model — the strictest tier, by design."	c3::ft02::analysis
"What does the Tulu 3 release (arXiv:2411.15124) add to the openness picture?"	"Tulu 3 (Allen Institute for AI) is the open-recipe POST-TRAINING stack — it releases the full multi-stage post-training recipe (data, code, configs), not just a base. It's the canonical 'open recipe for alignment/post-training' citation, complementing OLMo's open-base story."	c3::ft02::application
"Fill in the auditability column for these: OLMo-2, Llama-3.1-405B, GPT-4o."	"OLMo-2 = YES (open-recipe: data+code+checkpoints, Apache-2.0). Llama-3.1-405B = NO (weights-only: data described in aggregate, recipe summarized, Llama Community License). GPT-4o = NO (closed: weights withheld, API-only). Only OLMo-2 can answer 'what did it see?' from primary sources."	c3::ft02::application