"front"	"back"	"tags"
"Who is Ai2, and what is their canonical position in the open ecosystem?"	"Ai2 = Allen Institute for AI. The canonical FULLY-OPEN lab — they release weights + data + code + checkpoints + eval, all Apache-2.0. Their releases (OLMo, Tülu 3) are the strict end of the FT02 open spectrum — what 'open' means when a research lab commits to releasing everything."	c3::ftdd02::recall
"What are the two arXiv citations for OLMo 2 and Tülu 3?"	"OLMo 2: arXiv:2501.00656 (nicknamed '2 OLMo 2 Furious'). Tülu 3: arXiv:2411.15124 (the open post-training recipe). Together they are the canonical 'fully open' citations for the base and the post-training pipeline respectively."	c3::ftdd02::recall
"Name OLMo 3's three variants and the steer each represents."	"(1) base — pretrained weights, no post-training (Layer 1 only). (2) instruct — + SFT + DPO (Layer 3: format + preference). (3) think — + RLVR (Layer 3: + reasoning). Same base, three steers — the steer axis made concrete and directly comparable."	c3::ftdd02::recall
"Why is OLMo 3's three-variant release a 'pedagogical gift'?"	"It's the same base with three different steers (none / SFT+DPO / +RLVR). A student can load all three and directly SEE what each steering stage changed — diff base vs instruct (what SFT/DPO added), diff instruct vs think (what reasoning training added). A single-variant release cannot provide this. The clearest illustration of FT12/13/14."	c3::ftdd02::analysis
"What are the three stages of the Tülu 3 post-training pipeline?"	"(1) SFT — Supervised Fine-Tuning on curated instruction data (format, instruction-following). (2) DPO — Direct Preference Optimization on preference pairs (preference alignment). (3) RLVR — Reinforcement Learning on Verifiable Rewards (reasoning: math, code, logic). Each maps to Layer 3 of the FT00 stack; each stage's data/code/config is released."	c3::ftdd02::recall
"What is RLVR, and why is Tülu 3's implementation significant?"	"RLVR = Reinforcement Learning on Verifiable Rewards. The model generates solutions to problems with verifiable answers (math, code, logic), and correct solutions are rewarded. It's the technique behind the 'reasoning model' wave (FT14). Tülu 3 is significant because it is the OPEN implementation — the recipe is released, so you can study and reproduce it."	c3::ftdd02::application
"What is the open-instruct repo, and why does it matter?"	"allenai/open-instruct — the repository containing the Tülu 3 training code, data preparation scripts, and configuration files. It is the post-training analogue of OLMo's pretraining code: the executable recipe, not the description. If you want to reproduce Tülu 3 or build a Tülu-style pipeline on your own base, this repo is the primary source."	c3::ftdd02::recall
"State the three properties that only fully-open releases give you."	"(1) REPRODUCIBILITY (strict sense) — rebuild from pinned commit, prove no silent drift. (2) ABLATION — remove/change one component (data source, post-train stage) and measure the effect. (3) SUPPLY-CHAIN TRUST — audit pipeline end-to-end, rule out hidden training-time exfiltration. Each requires the full component stack (weights+data+code+checkpoints)."	c3::ftdd02::application
"Why is ablation impossible with a weights-only release?"	"Ablation requires removing or changing a single component (a data source, a post-training stage) and measuring the effect. With weights-only, the data and recipe are opaque — you can't 'skip RLVR' because you don't have the SFT/DPO/RLVR stages as separable components. Tülu 3 enables ablation because each stage is released separately and executable."	c3::ftdd02::analysis
"Contrast OpenBMB/MiniCPM and Ai2/OLMo-Tülu on orientation and model axis."	"OpenBMB/MiniCPM: PRODUCT/EDGE oriented; sizes 1B–4B + multimodal; variant axis = MODALITY (text/vision/omni); use = ship small model on edge. Ai2/OLMo-Tülu: RESEARCH oriented; sizes 7B–405B text-focused; variant axis = STEER (base/instruct/think); use = reproduce, audit, research. Both open-data/recipe; different centers of gravity."	c3::ftdd02::application
"What is the one-line heuristic contrasting MiniCPM and OLMo/Tülu?"	"'MiniCPM is what you SHIP. OLMo/Tülu is what you STUDY.' MiniCPM for small models on edge devices; OLMo/Tülu for understanding, auditing, and reproducing the full training/post-training stack (especially for regulated deployments)."	c3::ftdd02::analysis
"What does OLMo 2 release that establishes the fully-open standard?"	"Weights (multiple sizes, base + instruct) + the ACTUAL training corpus (not aggregate description) + training code/scripts + intermediate checkpoints (study training dynamics) + the evaluation suite (reproduce benchmarks). Everything a skilled person needs to rebuild the model from scratch."	c3::ftdd02::recall
"Why is 'treating fully-open as production-ready' an anti-pattern?"	"OLMo/Tülu are RESEARCH artifacts first — fully open, auditable, reproducible, but not always the most capable at their size. A closed or weights-only model may outperform them on benchmarks. Choose them for openness, auditability, reproducibility — NOT because they top leaderboards. Capability is FT03's concern."	c3::ftdd02::analysis
"Why is 'OLMo 3 think = GPT-class reasoning' a false inference?"	"The think variant is RLVR-tuned (the reasoning steer), but it's a 7B/32B model, not a frontier model. It demonstrates the RLVR technique OPENLY (its research contribution) but won't match a closed frontier reasoning model on the hardest tasks. Use it to UNDERSTAND reasoning fine-tuning (FT14), not to deploy a frontier reasoner."	c3::ftdd02::analysis
"How does Tülu 3's released pipeline enable the question 'what does RLVR actually contribute?'"	"Because the full post-training recipe is released (data+code+configs for SFT, DPO, and RLVR separately), a researcher can train Tülu 3 WITH and WITHOUT the RLVR stage, on the same OLMo base, with the same SFT and DPO, and directly measure RLVR's contribution. This ablation is impossible with a weights-only release where the stages aren't separable."	c3::ftdd02::application
"What are OLMo 3's two sizes and its release date?"	"7B and 32B. Released November 2025 (Nov 2025). All three variants (base/instruct/think) available at both sizes, all Apache-2.0."	c3::ftdd02::recall
"What is the maximum scale Tülu 3's pipeline has been run at?"	"Up to 405B parameters. Tülu 3's post-training recipe (SFT → DPO → RLVR) has been applied at scales up to 405B, demonstrating the pipeline is not just a small-model toy — it's a production-scale post-training methodology, released openly."	c3::ftdd02::recall
"Why does the course use BOTH MiniCPM and OLMo/Tülu?"	"Different roles: FT00's lab loads MiniCPM5-1B (cheap iteration — the on-ramp). FT02's audit references OLMo 2 (the fully-open standard). When teaching what post-training does (FT12 SFT, FT13 DPO, FT14 GRPO), OLMo 3's variants and Tülu 3's released pipeline are the clearest teaching tools. MiniCPM = iterate cheaply; OLMo/Tülu = understand deeply."	c3::ftdd02::analysis