Start from a research question and build a decision that can be checked later.
Choose a research mode: strict reproduction, method reproduction, modern rerun, or a new experiment.
Set the paper, role, update method, resources, access, and reproducibility constraints.
Read why candidates enter, what risks remain, and what still needs verification before comparing them.
Export the decision record with its task, candidates, comparisons, and evidence context.
Move from concepts to real models and research tasks.
Base is closer to continuation training; Instruct is tuned for instruction following; Thinking often adds explicit reasoning behavior or test-time compute. They are not a universal ranking, so reproduction must match the checkpoint type.
A checkpoint is a saved weight snapshot from a training stage. One family may have Base, Instruct, Thinking, and Coder checkpoints; reproduction must pin the checkpoint and revision, not only the family name.
Downloadable weights do not imply open training code, data, licensing, or a complete implementation. Check weights, base checkpoints, licenses, and derivative-distribution conditions separately.
API access lowers deployment friction but introduces version pinning, service availability, and data-boundary concerns. Local weights add hardware and licensing responsibilities. Callable does not mean reproducible.
LoRA usually updates adapters, SFT needs supervised data and a training recipe, and RL additionally needs an environment, reward, or evaluation loop. Inference suitability does not imply suitability for every update method.
Verified means the current record has sufficient supporting sources; partial means only some fields are evidenced. not_reported, not_disclosed, and not_verified are not false. Check sources and dates before strict reproduction.
Strict reproduction controls the original model and environment; method reproduction controls method variables while allowing comparable substitutes; modern reruns test whether conclusions transfer to current-generation models. Their success criteria differ.
Dense models use the same parameter set per token; MoE models activate only some experts. Total and active parameters are different quantities, so the site keeps them separate and shows missing values as unknown.
LoRA updates adapters, SFT updates weights through supervised training, and RL needs an environment, reward, and evaluation loop. External memory or prompt self-evolution is not weight updating; weight-updating self-evolution requires training and saving a new checkpoint.
Paper cases can expose three layers: a no-GPU learning track for the mechanism map, a reduced method check for the code and causal chain, and strict reproduction that aligns paper version, model revisions, data, baselines, hyperparameters, and compute. Their conclusions must remain separate.
Reproduction is more than a model name: pin the checkpoint, model revision, tokenizer, config, chat template, API version, and hardware conditions. When a paper does not report a field, keep not_reported or not_verified instead of filling a plausible value.
A saved weight snapshot at a training stage; Base, Instruct, and Thinking may be different checkpoints.
A note explaining what a source supports, what it does not support, and why an unknown state remains.
Pinning a model revision, commit, or API version so reproduction does not drift with service updates.
Hardware tiers are coarse research metadata, not an exact VRAM calculator.