id stringlengths 22 36 | topic stringclasses 20
values | topic_slug stringclasses 20
values | era stringclasses 3
values | facet stringlengths 79 224 | question stringlengths 90 353 | options dict | answer stringclasses 4
values | rationale stringlengths 225 803 |
|---|---|---|---|---|---|---|---|---|
reasoning_models_rlvr_0011 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Verifiable rewards: exact-match and symbolic checking for math, unit tests for code, and format rewards | In GRPO-style RLVR with binary verifiable rewards, what happens to the group-normalised advantage when every sampled rollout for a prompt receives the same reward? | {
"A": "The advantage equals the raw reward minus the running value-model baseline",
"B": "The advantage is zero for all rollouts, so the prompt yields no policy gradient",
"C": "The advantage is undefined, so the update falls back to the KL term only",
"D": "The advantage is clipped to the trust-region boundar... | B | Group normalisation subtracts the group mean, so an all-correct or all-wrong group has zero advantage and contributes no gradient, which is why dynamic sampling filters such groups out. Clipping applies to importance ratios, GRPO has no value model, and a zero advantage is well-defined rather than undefined. |
hardware_numerics_kernels_0239 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Kernel fusion: what fusing operations changes about memory traffic and launch overhead, and when hand-written kernels (CUDA/Triton) are warranted over library calls | A chain of five elementwise kernels, each already running at close to peak HBM bandwidth for several milliseconds, is rewritten as one fused Triton kernel. What quantity sets the speedup of the fused version over the chain? | {
"A": "The ratio of DRAM bytes moved by the chain to those moved by the fused kernel, since both remain bandwidth-limited",
"B": "The number of kernel launches removed, since each saved launch removes a fixed few microseconds of driver and dispatch cost",
"C": "The ratio of arithmetic intensity to the machine ba... | A | Both forms are bandwidth-limited, so time is proportional to DRAM bytes, and fusion's gain is exactly the traffic ratio from dropping intermediate stores and reloads. Launch overhead is a few microseconds against millisecond kernels, so removing four launches sets nothing. Fusion raises arithmetic intensity but pointwi... |
prompting_and_sft_0184 | Prompting and Supervised Fine-Tuning | prompting_and_sft | mixed | Sequence packing, attention masking between packed examples and their effect on training efficiency (the home for packing) | Some frameworks isolate packed segments without a custom kernel by adding a large negative bias to disallowed attention scores. Relative to a varlen (unpadded, cu_seqlens) kernel, what cost does this approach incur? | {
"A": "It doubles activation memory by retaining both the bias tensor and the unmasked scores for the backward pass",
"B": "It computes softmax scores for every cross-segment token pair and discards them once the bias is applied",
"C": "It forces the attention softmax into fp32 eager mode to keep the large negat... | B | The additive bias is applied after the full L×L score matrix has been evaluated, so all cross-segment pairs are computed and then thrown away, which is exactly the work a varlen kernel skips by iterating only over within-segment blocks. Position ids must be reset per segment in either scheme, so that is not a cost spec... |
prompting_and_sft_0141 | Prompting and Supervised Fine-Tuning | prompting_and_sft | mixed | Constructing tool-use demonstrations for SFT: call/result turns, masking of tool outputs, error-handling and recovery examples (the inference-side interface is under Agents) | What does interleaving the tool-use SFT mixture with general instruction-following data primarily protect against? | {
"A": "Exposure bias between teacher-forced training and autoregressive decoding",
"B": "Vocabulary drift caused by heavy exposure to JSON punctuation tokens",
"C": "Overfitting the loss mask to a single chat template variant",
"D": "Degradation of ordinary conversational ability from narrow-format training"
} | D | Training only on structured call traces narrows the output distribution and erodes plain chat quality, so mixtures keep general instruction data alongside tool data. Template overfitting is addressed by template variation, exposure bias is inherent to teacher forcing regardless of mixture, and token vocabularies do not... |
evaluation_methodology_0257 | Evaluation Methodology | evaluation_methodology | mixed | Multilingual and cultural evaluation: translated benchmarks versus natively built ones | Belebele covers 122 language variants with the same reading-comprehension items in every language. What property of its construction makes per-language accuracies directly comparable to one another? | {
"A": "Item difficulty was recalibrated per language with an item-response-theory model",
"B": "Passages were sampled from each language's own Wikipedia and news corpora",
"C": "Questions were authored independently by native speakers of each language variant",
"D": "Every passage, question and answer option i... | D | Belebele is fully parallel: it extends FLORES-200 passages, so all languages share identical content and any accuracy difference reflects the model, not the items. Independent authoring or per-language sourcing would break parallelism, and no IRT recalibration is applied. |
hardware_numerics_kernels_0113 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Mixed-precision training: master weights, loss scaling, and which operations must remain in higher precision | Dynamic loss scaling increases the scale factor after a run of successful steps. What failure mode would result from omitting this growth rule and keeping the initial scale fixed forever? | {
"A": "The optimiser would apply updates that are systematically too large by the scale factor",
"B": "The gradient norm used for clipping would drift away from its unscaled value",
"C": "Gradient magnitudes would shrink as training proceeds and increasingly underflow FP16",
"D": "Gradient magnitudes would gro... | C | Gradients typically shrink over training, so a scale chosen early becomes too small and small gradients flush to zero; periodic doubling tracks this. Overflow is what triggers backing the scale off rather than a consequence of a fixed scale, and the unscaling step keeps update magnitudes and clipping norms correct rega... |
transformer_architecture_0035 | Decoder-Only Transformer Architecture | transformer_architecture | fundamentals | Multi-query and grouped-query attention: KV-head sharing, memory savings at inference and quality trade-offs | In a decoder-only model whose decoding step is memory-bandwidth bound, moving from 32 KV heads to 4 shared KV heads reduces which per-token quantity roughly eightfold? | {
"A": "The bytes exchanged in the attention output all-reduce across tensor-parallel shards holding the replicated key-value heads",
"B": "The floating-point multiplies performed when each query head scores its group's keys over the full context window",
"C": "The parameters stored in the key and value projectio... | D | Each decoding step re-reads the whole KV cache, so cutting KV heads eightfold cuts per-token HBM read traffic by about eight and relieves the bandwidth bottleneck. The attention score multiplies are unchanged, since every query head still scores against its group's keys across the entire context. The shrinking of the k... |
hardware_numerics_kernels_0160 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Sub-8-bit number formats (MXFP4, NVFP4, INT4) as formats: block scaling, dynamic range and hardware support | In the OCP Microscaling MXFP4 format, each individual element is stored in a 4-bit floating-point encoding. Name the exponent/mantissa layout of that element format. | {
"A": "1 sign bit, 3 exponent bits, 0 mantissa bits (E3M0)",
"B": "1 sign bit, 2 exponent bits, 1 mantissa bit (E2M1)",
"C": "1 sign bit, 1 exponent bit, 2 mantissa bits (E1M2)",
"D": "0 sign bits, 2 exponent bits, 2 mantissa bits (E2M2)"
} | B | MXFP4 elements use the FP4 E2M1 encoding: sign, two exponent bits, one mantissa bit. E3M0 and E1M2 are conceivable 4-bit splits but are not the format the MX spec adopts; an unsigned E2M2 has no sign bit, which weights cannot tolerate. |
tokenisation_and_vocabularies_0301 | Tokenisation and Vocabularies | tokenisation_and_vocabularies | fundamentals | Alternatives to subword tokenisation: byte-level and character-level models, patching and dynamic tokenisation, and their costs | Hierarchical byte models such as Hourglass transformers downsample the sequence inside the network and must restore per-byte resolution before the output layer. Name the standard way the fine-grained information is preserved across the downsampling. | {
"A": "The pooled stream is upsampled by repeating each vector, and a shallow local-attention block over neighbouring bytes then sharpens the restored positions.",
"B": "The pre-downsampling byte activations are carried by a residual shortcut around the shrunken stack and added back after upsampling.",
"C": "Pat... | B | Hourglass-style stacks keep a skip path from the full-resolution activations and add it back after upsampling, so detail lost to pooling need not be reconstructed from the coarse stream. Upsampling by repetition plus local attention describes only how the coarse vectors are spread out, not how lost detail is retained; ... |
interpretability_in_practice_0074 | Interpretability in Practice | interpretability_in_practice | modern | Sparse autoencoders: dictionary learning on activations, sparsity penalties and feature interpretation | In sparse-autoencoder interpretability, a feature's label is causally tested by clamping that feature's activation to a chosen high or low value during the forward pass; what is then measured to complete the test? | {
"A": "The change in how often the dictionary's other features fire at the same layer afterwards",
"B": "The drop in accuracy of a linear probe for the labelled concept read from that layer",
"C": "The increase in the autoencoder's reconstruction error on the patched activation vector at that layer",
"D": "The... | D | The point of clamping is behavioural: one reads the resulting change in the model's own output distribution on label-relevant prompts, since only that connects the feature to what the model does. Reconstruction error measures the feature's contribution to the autoencoder's training objective rather than to model behavi... |
multimodal_language_models_0175 | Multimodal Language Models | multimodal_language_models | modern | Multimodal position encodings and how video and multi-image inputs are ordered and encoded | VideoRoPE and related analyses argue that allocating rotary dimensions so that the temporal axis rides on high-frequency components harms long-video retrieval. What allocation across the rotary dimensions do they recommend instead? | {
"A": "Assign temporal indexing to the low-frequency dimensions and the two spatial axes to high-frequency ones",
"B": "Assign temporal indexing to the high-frequency dimensions while giving the spatial axes a diagonally scaled index layout",
"C": "Interleave temporal and spatial indices in alternating dimension... | A | Low-frequency rotary dimensions have long wavelengths, so putting the frame index there gives monotone, non-aliasing order over hundreds of frames, while the short within-frame extents fit the high-frequency dimensions. Keeping the temporal index on high frequencies retains the periodic aliasing the work identifies, wh... |
interpretability_in_practice_0171 | Interpretability in Practice | interpretability_in_practice | modern | Model diffing: comparing representations across checkpoints and between base and fine-tuned models (stage-wise probes, crosscoders) to locate what training changed | In the "shallow safety alignment" analysis of safety-tuned chat models, per-token KL divergence between the aligned model and its base model, measured on harmful prompts, is concentrated at which positions of the generated sequence? | {
"A": "At the first few response tokens, where the aligned model's per-token distribution diverges most from the base.",
"B": "At the user-instruction tokens, where hidden states already diverge sharply before any response token is emitted.",
"C": "Evenly across the response, with per-token divergence from the b... | A | The measured divergence is sharply peaked at the opening few response tokens, which commit the model to a refusal or a compliance prefix — hence prefilling a compliant prefix largely undoes the tuning, and stage-wise diffs are most informative at those positions. Divergence does not stay flat across the response, is no... |
mixture_of_experts_0269 | Mixture-of-Experts Architectures | mixture_of_experts | modern | Upcycling dense checkpoints into MoE and other initialisation strategies for experts | Immediately after upcycling a dense checkpoint into an MoE with identical expert copies, what is the primary role of the auxiliary load-balancing loss during continued training? | {
"A": "To bound the magnitude of router logits so that the gating softmax stays numerically stable while training",
"B": "To equalise gradient norms reaching each expert so no duplicated copy drifts away faster than its siblings",
"C": "To inject noise into router logits so that the tied expert copies break symm... | D | With every expert computing the same function at initialisation, routing preferences are arbitrary and self-reinforcing, so the balancing term keeps token counts per expert near uniform until the copies differentiate. Constraining router logit magnitude for softmax stability is the separate router z-loss; adding noise ... |
prompting_and_sft_0104 | Prompting and Supervised Fine-Tuning | prompting_and_sft | mixed | Data quality versus quantity in SFT: small curated sets, filtering by model judges and diversity measures | A base LLM is instruction-tuned on 1,000 hand-curated examples; its MMLU accuracy stays close to the base model's while its human preference win rate rises sharply. What does this pattern indicate about the role of supervised fine-tuning? | {
"A": "It mainly limits catastrophic forgetting, so a small curated set retains pretrained facts that a hundred-thousand-example set would overwrite",
"B": "It mainly recalibrates output token probabilities, so gains show up in preference comparisons but never in likelihood-based evaluation scores",
"C": "It mai... | C | A thousand demonstrations cannot install substantial new world knowledge, so knowledge benchmarks move little while style, formatting and instruction adherence improve markedly — the elicitation view of SFT. The claim that new facts are hidden by multiple-choice scoring is wrong, since MMLU does register genuine knowle... |
pretraining_data_pipelines_0122 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Educational-value and domain classifiers for filtering (FineWeb-Edu and DCLM style) and the risk of narrowing the distribution | Educational-value classifiers of the FineWeb-Edu and DCLM kind are usually applied to English web crawl text but not to the curated math or code sources in the same pretraining mixture. What is the reason for exempting those curated sources from the scorer? | {
"A": "DCLM's fastText filter uses instruction data and ELI5 answers as positives against sampled crawl, which lifts the share of explanatory prose retained from the pool",
"B": "Aggressive educational thresholds discard most of the crawl pool, so token budgets at larger compute scales must be met by repeating the... | C | The educational-value scorer is calibrated on annotated web prose, so its output on code or LaTeX measures unfamiliarity with that register rather than usefulness, and those sources are quality-controlled upstream instead. The claim about a regression head over embeddings at a score-three cut describes how the FineWeb-... |
safety_alignment_practice_0009 | Safety and Alignment Practice | safety_alignment_practice | mixed | Refusal training: harmful-request datasets, borderline cases and measuring over-refusal against under-refusal | Safe-RLHF separates the two objectives that ordinary RLHF blends into one scalar. State how the separation is implemented during optimisation. | {
"A": "A single reward model is trained on annotations that mark helpfulness and harmlessness with separate rank labels",
"B": "A harmlessness classifier vetoes rollouts before they reach the policy-gradient update, leaving the reward untouched",
"C": "Two reward models are averaged with a fixed weight tuned onc... | D | Safe-RLHF trains a distinct cost model for harm and solves a constrained problem where the multiplier on the cost is updated during training. Fixed-weight averaging, a single multi-label reward model and a rollout veto are plausible alternative designs but not the method's mechanism. |
scaling_laws_compute_allocation_0270 | Scaling Laws and Compute Allocation | scaling_laws_compute_allocation | fundamentals | Scaling laws for specialised settings: MoE sparsity, vocabulary size and distillation as additional axes, and how numeric precision enters loss fits (the formats themselves are under Hardware) | The precision scaling laws predict a different answer when model size is held fixed rather than optimised. In that constrained case, how does the optimal training precision change as the training token budget grows? | {
"A": "It decreases, because more data compensates for the capacity lost to coarse weight quantization",
"B": "It increases, because a fixed-size model must represent more information learned from more data",
"C": "It decreases until the token-to-parameter ratio reaches roughly 100 and then increases again",
"... | B | With N pinned, extra data pushes the model to use its limited capacity harder, so the optimum shifts toward higher bit-width — the paper notes this can grow faster than proportionally in compute. The other trends contradict the constrained-optimisation result, though each mirrors an intuition someone might hold. |
distributed_training_systems_0020 | Distributed Training Systems | distributed_training_systems | fundamentals | Data parallelism and gradient all-reduce: ring and tree algorithms, bandwidth cost and overlap with backward computation | NCCL's double binary tree algorithm uses two complementary trees over the same ranks instead of one. Name the benefit that the second tree provides. | {
"A": "It provides a redundant copy of the gradients, tolerating a single rank failure",
"B": "It permits the reduction to run in fp32 while the transfer remains in bf16",
"C": "It halves the depth of the reduction, cutting latency steps to log4(N)",
"D": "It lets every rank act as an interior node, so all lin... | D | In a single binary tree half the ranks are leaves with idle upstream bandwidth; the complementary tree makes leaves into interior nodes so full duplex bandwidth is exploited. Depth stays logarithmic base two, the scheme offers no fault tolerance, and reduction dtype is an orthogonal NCCL setting. |
agents_and_tool_use_0184 | Agents and Tool Use | agents_and_tool_use | modern | Multi-step RL for agents: credit assignment across turns, trajectory-level rewards and sample cost | Generalised advantage estimation with λ close to 1 is applied to a long agent trajectory. Relative to a small λ, what is the effect on the advantage estimates for early turns? | {
"A": "Lower variance and higher bias, since estimates lean on the critic's value predictions",
"B": "Higher variance and lower bias, since estimates lean on the realised trajectory return",
"C": "Unchanged bias and variance, since λ only rescales the estimates by a constant factor",
"D": "Lower variance and l... | B | λ→1 approaches the Monte-Carlo return: nearly unbiased but high variance, and this trade-off is most severe for early turns of a long episode. The low-variance/high-bias regime is λ→0; λ genuinely alters the bias-variance mix rather than rescaling; and no setting of λ improves both simultaneously. |
reasoning_models_rlvr_0088 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Emergent long chain-of-thought behaviours: self-verification, backtracking and the growth of response length during RL | Some RLVR runs show accuracy rising while average length briefly falls, before length resumes growing. What is the usual explanation for the early length drop? | {
"A": "Advantage normalisation by token count penalises long sequences early on",
"B": "The policy first prunes rambling, off-task text inherited from the base model",
"C": "The KL penalty pulls the policy back toward the shorter reference distribution",
"D": "The entropy bonus is annealed down, shortening sam... | B | Early RL cleans up unfocused base-model verbosity, which shortens traces before genuine search behaviour lengthens them again. KL pull-back, entropy annealing and token-count normalisation are real mechanisms in these pipelines but do not account for the characteristic early dip. |
agents_and_tool_use_0043 | Agents and Tool Use | agents_and_tool_use | modern | Agent loops: ReAct-style interleaving of reasoning and actions, scratchpads and termination conditions | In a standard function-calling agent loop, the model returns an assistant turn whose content is plain text with no tool call. What does the harness conventionally do with that turn? | {
"A": "Append it to the scratchpad as an observation and issue another model call for the next step",
"B": "Treat it as a malformed action and re-prompt the model with a reminder of the tool schema",
"C": "End the loop and return that text to the caller as the agent's final answer",
"D": "Retain it in the tran... | C | The absence of tool calls on an assistant turn is the standard stop condition: the harness exits the loop and surfaces the content as the final answer. Appending it as an observation and re-invoking the model is how genuine tool results are handled, not final text. Re-prompting with the schema is the response to an unp... |
optimisers_training_stability_0278 | Optimisers and Training Stability | optimisers_training_stability | fundamentals | Checkpoint averaging, exponential moving averages of weights and their effect on final quality | WiSE-FT linearly interpolates the weights of a fine-tuned CLIP model with the original zero-shot weights. What is the main empirical gain it reports over the fine-tuned model alone? | {
"A": "Large improvements in in-distribution accuracy, with zero-shot accuracy fully restored too",
"B": "Large reductions in calibration error measured on the fine-tuning distribution itself",
"C": "Large reductions in fine-tuning compute, since far fewer epochs are needed to converge",
"D": "Large improvemen... | D | WiSE-FT's headline result is robustness: interpolation recovers much of the pretrained model's out-of-distribution accuracy while matching or exceeding fine-tuned in-distribution accuracy. It does not shorten fine-tuning, does not restore zero-shot behaviour exactly, and calibration on the fine-tuning distribution is n... |
distributed_training_systems_0074 | Distributed Training Systems | distributed_training_systems | fundamentals | Tensor parallelism: splitting attention heads and MLP matrices, the all-reduce pattern per layer and why it stays within a node | In plain tensor parallelism (Megatron-style) without sequence parallelism, how are a transformer layer's LayerNorm weight and bias parameters handled across the tensor-parallel ranks? | {
"A": "They live on rank zero in fp32 master form and are broadcast to the other ranks at the start of every forward pass.",
"B": "Every rank holds an identical copy and redundantly normalises the same full-width activation, so no collective is needed for LayerNorm.",
"C": "Each rank normalises a distinct slice ... | B | In plain tensor parallelism the layer's input activation is already full-width on every rank (the previous all-reduce restored it), so each rank keeps an identical copy of the small normalisation parameters and repeats the same computation, costing duplicated FLOPs but no communication. Splitting the region along token... |
tokenisation_and_vocabularies_0013 | Tokenisation and Vocabularies | tokenisation_and_vocabularies | fundamentals | Byte-pair encoding: merge training, frequency-based vocabulary growth, and why byte-level fallback avoids unknown tokens | In a GPT-2 style byte-level BPE tokeniser, a leading space is normally absorbed into the token for the word that follows. If a prompt ends with an explicit trailing space, which part of the vocabulary must supply the next generated token, and with what effect on completion quality? | {
"A": "The next token must be drawn from the single-byte tokens reserved for fallback, since a standalone space token can only be followed by raw bytes, which fragments the word and slows decoding.",
"B": "The next token must be drawn from the same space-prefixed tokens as usual, because the tokeniser strips trail... | D | Because merges attach the space to the following word, a bare trailing space is tokenised alone and the model must continue with the space-less word-initial pieces, which are far rarer in training and give worse completions. Byte fallback is triggered only by byte sequences absent from the merge table, not by a precedi... |
synthetic_data_and_distillation_0105 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Model collapse, distribution narrowing and stylistic fingerprints from recursive training on generated data, including identity leakage as a symptom | A team measures collapse by tracking perplexity of each generation's model on a held-out corpus of original human text. What trend indicates collapse is underway? | {
"A": "Perplexity on the model's own samples rises across generations",
"B": "Perplexity on the held-out human text stays flat across generations",
"C": "Perplexity on the model's own samples matches held-out human text",
"D": "Perplexity on the held-out human text rises across generations"
} | D | As the model narrows onto its own output distribution, real human text — which contains the discarded tail — becomes less likely under it, so held-out perplexity climbs. Perplexity on self-samples typically falls rather than rises during collapse; flat held-out perplexity indicates no collapse; matching perplexities is... |
evaluation_methodology_0312 | Evaluation Methodology | evaluation_methodology | mixed | Evaluating instruction following and formatting with verifiable constraints, and coverage gaps in such tests | IFEval-style benchmarks check constraints such as "exactly 200 words" or "avoid the letter e" with deterministic Python verifiers. Name the capability an LLM judge would have to possess to replace those verifiers on such counting constraints, which the tested models themselves demonstrably lack. | {
"A": "The judge would have to suppress self-preference, since judges score responses from their own pretraining family higher than equivalent outputs from unrelated model families.",
"B": "The judge would have to count words and characters exactly, the same operation the tested models fail at, so its grading erro... | B | A deterministic verifier counts exactly; an LLM judge grading a length or letter-frequency constraint must perform that very counting, which language models do unreliably, so judge error is correlated with the constraints the benchmark targets. Position sensitivity and self-preference are genuine judge pathologies, but... |
hardware_numerics_kernels_0093 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Floating-point formats: FP32, TF32, BF16, FP16 and their exponent/mantissa trade-offs for training | In ring all-reduce over N data-parallel workers, each chunk is added into a running partial sum once per hop of the reduce-scatter phase. If those additions are done in BF16 instead of being accumulated in FP32, how does the rounding error on a gradient element behave as N grows? | {
"A": "It grows only as the square root of N, since hop order varies between runs and the per-hop roundings act as independent zero-mean perturbations.",
"B": "It accumulates over the N-1 sequential additions, so the worst-case relative error grows roughly linearly with worker count, which FP32 accumulation bounds... | B | In reduce-scatter a given element is summed once at each of N-1 hops, so with only 7 mantissa bits the rounding errors chain and the worst-case relative error scales with the number of hops; keeping the running sum in FP32 bounds it. The claim that a single division rounding dominates is wrong because that division is ... |
open_model_landscape_0062 | Open Model Landscape Since 2024 | open_model_landscape | modern | Licence and terms constraints on open weights and on distilling from proprietary models: permissive, community and use-restricted licences and their practical implications for derivatives and synthetic data | Hugging Face hosts many models under gated access requiring users to accept terms before download. What legal function does the click-through gate serve for the model publisher? | {
"A": "It transfers copyright in derivative weights back to the publisher",
"B": "It converts the licence into a contract enforceable in the EU only",
"C": "It obliges the hub to indemnify the publisher against downstream misuse",
"D": "It records evidence that each downloader assented to the licence terms"
} | D | The gate creates an assent record, strengthening the publisher's position that recipients accepted the licence conditions rather than merely receiving a copy. It effects no copyright assignment, does not limit enforceability to any single jurisdiction, and creates no indemnity obligation on the hosting platform. |
open_model_landscape_0246 | Open Model Landscape Since 2024 | open_model_landscape | modern | Multilingual and region-specific open models: the training decisions behind them (tokeniser design, language mixture, translated post-training data) and their trade-offs against general models | Compared with the equivalent English text, generating a fixed amount of Bengali text on a model with poor Bengali segmentation increases which serving cost most directly? | {
"A": "Scheduler overhead, since more requests must be batched per second",
"B": "KV-cache memory per request, since more tokens are cached per output",
"C": "Weight memory, since more embedding rows are touched per sequence",
"D": "Prefill FLOPs only, since decoding cost is independent of token count"
} | B | High fertility means more tokens per response, and both decode steps and KV-cache footprint scale with token count, so per-request cache memory grows directly. Weight memory is fixed regardless of which rows are read, decoding cost is very much proportional to token count, and the number of requests does not change. |
multimodal_language_models_0186 | Multimodal Language Models | multimodal_language_models | modern | Multimodal position encodings and how video and multi-image inputs are ordered and encoded | In a decoder-only VLM using ordinary 1-D RoPE, a 1024-patch image is placed before a long text answer. Compared with M-RoPE, what does this cost in terms of the model's positional budget? | {
"A": "It consumes 1024 position ids instead of one shared id per image",
"B": "It requires interpolating rotary frequencies by a factor of 32",
"C": "It requires a separate attention mask over the visual span",
"D": "It consumes 1024 position ids instead of about the grid's side length"
} | D | Flattening spends one id per patch, whereas the factorised scheme advances only by the maximum of the height and width extents, roughly 32 for a 32x32 grid. Collapsing an image to a single id is not what M-RoPE does, frequency interpolation is a separate long-context trick, and neither scheme needs an extra visual mask... |
reasoning_models_rlvr_0092 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Emergent long chain-of-thought behaviours: self-verification, backtracking and the growth of response length during RL | Several groups report that a reasoning model's self-verification step, when it flips an answer, changes it in a measurable direction on easy benchmark items. What is that direction? | {
"A": "Flips occur only on items where the first answer failed format parsing",
"B": "Flips are roughly balanced, leaving accuracy on easy items unchanged",
"C": "Flips more often turn a wrong answer into a correct one, raising accuracy",
"D": "Flips more often turn a correct answer into a wrong one, lowering ... | D | On easy items the first answer is usually already right, so continued second-guessing is net harmful — the core of the overthinking finding. Balanced flips, net-corrective flips and format-triggered flips are each defensible-sounding but contradict the reported asymmetry. |
tokenisation_and_vocabularies_0303 | Tokenisation and Vocabularies | tokenisation_and_vocabularies | fundamentals | Alternatives to subword tokenisation: byte-level and character-level models, patching and dynamic tokenisation, and their costs | For a dynamic-patching byte-level model (BLT-style) to remain usable for autoregressive generation, the patch-boundary module is constrained in one specific way. State that constraint. | {
"A": "It must be causal, deciding each boundary from preceding bytes only, so segmentation can proceed as bytes are emitted.",
"B": "It must be deterministic across runs, so that identical documents yield identical patches whatever the batch size or hardware used.",
"C": "It must be differentiable, so that boun... | A | At generation time the following bytes do not yet exist, so the boundary decision at each position can only condition on the prefix; entropy-based patching uses a small causal byte-level LM for exactly this reason. Reproducibility across batch sizes and capping patch length are engineering conveniences that neither ena... |
long_context_and_retrieval_0125 | Long Context and Retrieval | long_context_and_retrieval | modern | Linear attention and state-space alternatives (Mamba-style) and hybrid transformer–SSM stacks for long sequences | A Mamba-style SSM compresses all history into a fixed-size recurrent state. In hybrid transformer–SSM stacks, what capability do the few interleaved attention layers supply that the SSM layers cannot provide as context length grows? | {
"A": "Exact retrieval of an arbitrary earlier token, since attention retains a per-token key–value cache whose size grows with the sequence rather than a fixed-width summary of the past.",
"B": "Input-dependent gating of the state transition, so that irrelevant tokens are attenuated and salient ones are written i... | A | Attention keeps a growing KV cache, so it can copy or look up any specific earlier token exactly — the associative-recall ability a constant-size recurrent state loses as context grows, which is precisely why hybrids add a few attention layers. Content-selective gating is exactly what Mamba's selection mechanism alread... |
hardware_numerics_kernels_0012 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Accelerator memory hierarchy: HBM, on-chip SRAM/shared memory and registers, and bandwidth versus capacity at each level | A CUDA kernel's occupancy drops when each thread block requests a large shared-memory allocation. What is the mechanism behind this drop? | {
"A": "Shared memory per SM is a fixed capacity, so fewer blocks can be simultaneously resident on a single SM",
"B": "The register file per SM is finite, so a high per-thread register count caps the number of resident warps",
"C": "Shared memory and L1 cache occupy one physical SRAM array whose split is configu... | A | Occupancy is set by whichever per-SM resource binds first; a large per-block shared-memory footprint against a fixed per-SM SRAM capacity limits how many blocks can be co-resident, so fewer warps are available to hide latency. Bank conflicts serialize accesses and reduce shared-memory throughput but do not change block... |
transformer_architecture_0301 | Decoder-Only Transformer Architecture | transformer_architecture | fundamentals | Encoder-decoder and encoder-only models compared with decoder-only, and where each remains in use | UL2 pretrains a single model on several denoising configurations that differ in corruption rate and mean span length, one of which is causal-style. Name this pretraining objective. | {
"A": "Fill-in-the-middle blended with causal modelling at a fixed rate, delimited by prefix and suffix sentinels",
"B": "Span corruption applied at several corruption rates, with a distinct sentinel token marking each masked span",
"C": "Mixture-of-denoisers, sampled per example and signalled to the model by a ... | C | UL2's objective is named the mixture-of-denoisers: R-, S- and X-denoising configurations are sampled per example and a paradigm token tells the model which mode is in force, allowing mode switching at inference. Multi-rate span corruption with sentinels describes T5-style masking without the paradigm signalling or the ... |
open_model_landscape_0290 | Open Model Landscape Since 2024 | open_model_landscape | modern | Choosing an open model for a use case: matching size, licence, context length, modality and post-training to constraints, and validating the choice with task-specific evaluations | A team must run an open model on a single 24 GB GPU with a long-context chat workload. Which architectural detail of the candidate release most strongly determines whether long conversations fit in memory alongside the weights? | {
"A": "The vocabulary size, which sets the size of the embedding and output projection matrices",
"B": "The number of key-value heads, which sets the per-token KV cache footprint",
"C": "The MLP expansion ratio, which dominates the per-layer parameter count",
"D": "The use of tied input and output embeddings, ... | B | KV cache size scales with sequence length times layers times KV heads times head dimension, so grouped/multi-query attention with few KV heads is what makes long contexts fit. Vocabulary size, MLP ratio and embedding tying affect the fixed weight footprint, which does not grow with conversation length. |
multimodal_language_models_0272 | Multimodal Language Models | multimodal_language_models | modern | Multimodal evaluation: document and chart understanding, visual reasoning benchmarks and hallucination of visual content | TextVQA differs from DocVQA in the kind of images it uses for reading text. What image source does TextVQA draw on? | {
"A": "Scanned industry documents such as reports, forms and typed letters",
"B": "Natural photographs of everyday scenes containing incidental written text",
"C": "Rendered synthetic charts paired with template-generated numeric questions",
"D": "Screenshots of mobile application interfaces annotated with wid... | B | TextVQA is built on Open Images photographs where signs, labels and packaging carry the text to be read. Scanned industry documents describe DocVQA's source, synthetic rendered charts describe chart-QA construction, and app screenshots describe GUI benchmarks such as ScreenSpot. |
interpretability_in_practice_0197 | Interpretability in Practice | interpretability_in_practice | modern | Reading intermediate representations: logit lens and tuned lens, early decoding, and layer-wise probing of what is predicted where | A linear probe trained on layer-12 activations classifies a sentence property with 95% accuracy. Considered alone, what does this result establish? | {
"A": "That the property is linearly decodable from that layer's activations by some external classifier.",
"B": "That the property is written into the residual stream by the attention heads feeding that layer.",
"C": "That the property is causally used by the model when it produces its output on those inputs.",... | A | Probe accuracy is a statement about decodability by the probe, nothing more; causal use requires intervention experiments such as steering or ablation. Single-direction encoding, causal use, and attribution to particular writing components are all further claims that probe accuracy alone cannot support. |
evaluation_methodology_0174 | Evaluation Methodology | evaluation_methodology | mixed | Statistical rigour: confidence intervals, sample sizes, variance across seeds and reporting of multiple runs | A generative benchmark of N items is scored by drawing k samples per item and averaging the per-item pass rates. Ignoring model-side correlation between draws, how does the standard error of the overall score scale? | {
"A": "As one over the square root of k alone, with N fixed by the benchmark",
"B": "As one over N multiplied by the square root of k, combining both counts",
"C": "As one over the square root of N alone, because averaging the k draws within each item removes within-item sampling noise before the items are poole... | D | Under the stated idealisation every draw is an independent Bernoulli trial, so the overall mean rests on Nk independent draws and its standard error falls as one over the square root of Nk. Scaling with k alone or with N alone each drops one of the two dimensions; the claim that within-item averaging removes sampling n... |
inference_and_serving_0293 | Inference and Serving Systems | inference_and_serving | mixed | Serving metrics: time to first token, inter-token latency, goodput, and cost per token under different batch sizes | For a request with a P-token prompt and N generated tokens, write how end-to-end latency is normally decomposed in terms of the two commonly reported serving metrics. | {
"A": "TTFT plus N times the mean inter-token latency",
"B": "TTFT plus (N minus one) times the mean inter-token latency",
"C": "P times the mean inter-token latency plus TTFT",
"D": "TTFT times N divided by the mean inter-token latency"
} | B | TTFT already covers prefill and the first emitted token, so the remaining N-1 tokens each cost one inter-token interval. Multiplying by N double-counts the first token. Scaling by the prompt length P confuses prefill cost with decode intervals. The last expression is not dimensionally a latency. |
multimodal_language_models_0083 | Multimodal Language Models | multimodal_language_models | modern | Native early-fusion multimodality: shared token spaces, discrete image tokens and single-model training from scratch | Chameleon reordered its normalisation layers away from the standard pre-norm placement, rather than using QK-norm or a z-loss, to bound one specific quantity that grew during mixed-modality training. What quantity did that reordering bound? | {
"A": "The variance of the embedding-table gradients differing between the modalities",
"B": "The magnitude of the attention logits entering the per-head softmax",
"C": "The norm of the residual stream activations accumulated across depth",
"D": "The magnitude of the output softmax normalising constant over th... | C | With pre-norm, each block writes an unnormalised contribution into the residual stream, so its norm grows with depth; modality-specific scale differences amplified this, and Chameleon adopted Swin-style norm reordering (normalising block outputs) to bound it. Attention-logit growth was instead handled by query-key norm... |
multimodal_language_models_0271 | Multimodal Language Models | multimodal_language_models | modern | Multimodal evaluation: document and chart understanding, visual reasoning benchmarks and hallucination of visual content | In long free-form image descriptions generated by vision-language models, how does the measured rate of object hallucination change with position in the output? | {
"A": "It grows across later sentences, because decoding leans increasingly on text co-occurrence priors",
"B": "It stays roughly constant, because attention over the same image tokens is recomputed at every step",
"C": "It peaks in the opening sentence, where the model commits to a global scene category before ... | A | Caption-level studies (e.g. CHAIR computed over sentence position) find fabricated objects concentrate near the end of long descriptions, where the decoder increasingly follows language-model co-occurrence statistics rather than visual evidence. A declining rate would follow if early grounding constrained later text, b... |
reasoning_models_rlvr_0254 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Running RL directly on small models versus SFT on teacher traces: which reasoning behaviours transfer, why small models often fail under direct RL, and the compute comparison (distillation mechanics are under Synthetic Data) | Reward shaping for code RLVR on a weak small model often replaces a single all-tests-pass bit with a fraction-of-tests-passed score. What problem specific to weak policies does that change address? | {
"A": "It removes the need to run tests in a sandbox, since partial output is scored statically",
"B": "It provides non-zero reward variance on tasks the model never solves completely",
"C": "It penalises overlong solutions, since more tests imply more required code",
"D": "It prevents reward hacking by models... | B | Dense partial credit gives differing rewards within a group even when no rollout fully passes, restoring a gradient where a binary reward would be uniformly zero. Partial credit actually makes test-special-casing hacking easier, sandboxing is still required, and it says nothing about solution length. |
tokenisation_and_vocabularies_0152 | Tokenisation and Vocabularies | tokenisation_and_vocabularies | fundamentals | Tokenisation fertility across languages and scripts and its effect on context budget, cost and quality for non-English text | Thai is written without spaces between words. What consequence does this have for a BPE tokeniser whose pre-tokenisation regex splits on whitespace before learning merges? | {
"A": "Long space-free spans are treated as single pre-tokens, so merges must cover whole clauses",
"B": "Every Thai character becomes its own pre-token, since the regex finds no word boundaries at all",
"C": "Thai text is routed to a byte-fallback path that bypasses the learned merge table entirely",
"D": "Me... | A | Whitespace pre-tokenisation cannot cut inside a space-free Thai run, so BPE receives long, mostly unique spans that merges must chew through, raising fertility. A whitespace regex does not emit one pre-token per character; byte fallback is triggered by unseen bytes, not by script identity; and standard GPT-style regexe... |
prompting_and_sft_0002 | Prompting and Supervised Fine-Tuning | prompting_and_sft | mixed | Chat templates and system prompts: role markers, template mismatch between training and inference and its symptoms, and the persona and formatting behaviours learned from the system turn | During SFT on chat-formatted data, the end-of-turn token that closes the assistant message is left out of the supervised (loss-bearing) span. What symptom does this produce at inference? | {
"A": "Generation ignores the system-turn persona and reverts to a neutral default voice, since role markers lose their conditioning effect without a supervised terminator",
"B": "Generation stops after only a few tokens, because loss on the prompt span dominates and pushes probability mass toward the terminator e... | C | Stopping is a learned prediction: if the turn terminator never carries loss, the model assigns it little probability and keeps sampling, typically rolling on into the next role marker and hallucinating the user's reply. Premature stopping arises from the opposite condition (terminators over-represented or short truncat... |
synthetic_data_and_distillation_0138 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Filtering synthetic data: model judges, verifiers, consistency checks and deduplication of generated samples | A verifier that accepts an incorrect solution is a false positive, and one that rejects a correct solution is a false negative. For building a synthetic SFT set by rejection sampling, state which error type is the more damaging and why. | {
"A": "False negatives, because they bias the retained samples toward shorter completions",
"B": "False positives, because they inflate the measured pass rate used for checkpoint selection",
"C": "False negatives, because they shrink the dataset and remove the hardest usable items",
"D": "False positives, beca... | D | Accepted-but-wrong samples enter the gradient and teach the error, whereas rejecting a correct sample only wastes compute and data. The false-negative options describe real but milder costs, and the inflated-pass-rate claim is an evaluation problem rather than the primary training harm. |
inference_and_serving_0032 | Inference and Serving Systems | inference_and_serving | mixed | KV cache mechanics and memory sizing, paged/blocked allocation, and prefix caching for shared-prompt reuse across requests | For a transformer with 32 layers, 8 key-value heads and head dimension 128, served in fp16, what is the KV-cache footprint of a single token across the whole model? | {
"A": "512 KiB",
"B": "64 KiB",
"C": "128 KiB",
"D": "256 KiB"
} | C | Per token: 2 (key and value) x 32 layers x 8 heads x 128 dims x 2 bytes = 131072 bytes = 128 KiB. 64 KiB drops either the key/value factor or the fp16 byte count; 256 KiB and 512 KiB overcount by assuming 16 heads or fp32 storage. |
synthetic_data_and_distillation_0063 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Logit distillation: soft-target KL losses, temperature and on-policy distillation from student samples | A student is distilled with the loss KL(teacher || student) computed on tokens sampled from the teacher. If the student later encounters a prefix that the teacher would essentially never produce, what does the training objective say about the student's behaviour there? | {
"A": "It is tightly constrained, because the mass-covering loss forces agreement over the whole input space",
"B": "It is regularised by the T squared factor, which uniformly scales gradients across all visited prefixes",
"C": "It is bounded by the reverse KL, which penalises any student mass outside the teache... | D | Forward KL is an expectation under the teacher's own distribution, so states the teacher never visits contribute almost nothing and the student's behaviour there is untrained, which is precisely the compounding-error motivation for on-policy distillation. Mass covering applies over the teacher's next-token support at v... |
synthetic_data_and_distillation_0262 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Pruning-plus-distillation compression: depth and width pruning of large checkpoints followed by distillation, versus training small models from scratch | After pruning, Sheared LLaMA continues pretraining with 'dynamic batch loading'. What does this procedure adjust during training? | {
"A": "The mixing weight of the distillation and language-modelling terms in the loss",
"B": "The sequence length of packed examples, according to remaining compute budget",
"C": "The sampling proportions of the pretraining domains, according to per-domain loss progress",
"D": "The global batch size, according... | C | Dynamic batch loading reweights domain sampling so that domains whose loss has not yet reached its reference level get more data, correcting the uneven damage pruning does across domains. Batch-size schedules, loss-term mixing and sequence-length curricula are all real training knobs but not what this method changes. |
inference_and_serving_0307 | Inference and Serving Systems | inference_and_serving | mixed | Serving metrics: time to first token, inter-token latency, goodput, and cost per token under different batch sizes | A RAG service prepends a long fixed context that is identical across all requests, and the serving engine enables prefix caching. State which serving metric this most directly reduces and by what mechanism. | {
"A": "Goodput, because the scheduler can admit requests whose prefix already resides in the block pool ahead of others",
"B": "Time to first token, because prefill reuses the already-computed key-value entries and processes only the new suffix",
"C": "Inter-token latency, by shortening the span of key-value ent... | B | Prefix caching is a prefill optimization: the shared prefix's key-value entries are reused, so only the novel suffix is computed and the wait before the first token collapses. Decode still attends over the full cached context, so the per-step attention read is unchanged. Sharing blocks does save memory and can indirect... |
evaluation_methodology_0110 | Evaluation Methodology | evaluation_methodology | mixed | LLM-as-judge evaluation: pairwise judging, position and verbosity biases, and calibration against human ratings | Chatbot Arena samples which pair of models to show a user non-uniformly. What is the stated purpose of this adaptive pair-sampling scheme? | {
"A": "To balance how often each model is placed in the first presented slot, cancelling out position bias",
"B": "To keep the comparison graph connected so the Bradley-Terry likelihood retains a unique maximiser over models",
"C": "To identify voters whose judgements diverge from crowd consensus so their ballot... | D | Adaptive pair selection concentrates the limited vote budget on comparisons that are most informative about uncertain rating differences, tightening confidence intervals faster than uniform sampling. Randomising presentation order handles position bias, but that is a separate mechanism from which pair is chosen; connec... |
safety_alignment_practice_0209 | Safety and Alignment Practice | safety_alignment_practice | mixed | Memorisation and privacy: training-data extraction, membership inference, PII leakage, and unlearning methods and how their success is evaluated | Carlini et al.'s 'Quantifying Memorization Across Neural Language Models' measured how much verbatim training text a model emits. State the functional form they found relating memorisation to model scale, sequence duplication count, and prompt context length. | {
"A": "Memorisation grows quadratically in model size but is independent of how often a sequence is duplicated",
"B": "Memorisation is a step function of duplication count, with a sharp threshold near ten copies",
"C": "Memorisation falls off exponentially once the number of training epochs exceeds one",
"D": ... | D | The paper's central empirical result is a log-linear scaling of memorisation with model capacity, duplication count and context length. Duplication clearly matters, so the quadratic-and-independent claim is wrong; no sharp ten-copy threshold or exponential epoch decay was reported, though duplication and repeated expos... |
interpretability_in_practice_0082 | Interpretability in Practice | interpretability_in_practice | modern | Sparse autoencoders: dictionary learning on activations, sparsity penalties and feature interpretation | In a sparse autoencoder trained on model activations with an L1 penalty, decoder columns are constrained to unit norm (or their gradients projected to preserve that norm). What degenerate solution does this constraint block? | {
"A": "Learning two dictionary columns pointing in the same direction, so a single concept splits its activation across both",
"B": "Shrinking feature activations toward zero while enlarging decoder columns proportionally, leaving the reconstruction unchanged",
"C": "Driving the encoder bias so negative that a l... | B | Without a norm constraint the L1 term can be made arbitrarily small by scaling activations down and decoder column norms up by the same factor, since the product reconstructing the input is unchanged; unit-norm columns remove this reparameterisation. Duplicated decoder directions (feature splitting), dead features caus... |
pretraining_data_pipelines_0241 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Code and mathematics corpora: sourcing from repositories and web, licence and quality filtering, and effects on reasoning | The Stack's licence detection stage relies on a specific automated tool applied to repository files rather than on GitHub's reported licence field. Name the tool used. | {
"A": "Semgrep, which matches rule patterns against a project's source tree",
"B": "ScanCode, which detects licence text and SPDX identifiers in files",
"C": "Trufflehog, which scans repository contents for embedded credentials",
"D": "Tree-sitter, which parses files to confirm they are valid source code"
} | B | ScanCode (with go-license-detector in early versions) is the licence scanner used to assign permissive licences to repositories in The Stack. Trufflehog is a secrets scanner, tree-sitter is a parser used for syntax filtering, and Semgrep is a static-analysis rule engine; none performs licence identification. |
safety_alignment_practice_0261 | Safety and Alignment Practice | safety_alignment_practice | mixed | Fine-tuning attacks on open weights and the fragility of safety training under further training | Studies of post-training compression report a safety-relevant side effect when an aligned chat model is aggressively quantised or pruned without any adversarial data. What happens to its refusal behaviour? | {
"A": "It weakens measurably, raising compliance with harmful prompts",
"B": "It stays intact because refusal is robust to low-precision weight rounding",
"C": "It becomes stronger, since degraded fluency makes harmful completions incoherent",
"D": "It shifts to depending on the calibration set used for the co... | A | Multiple evaluations find that quantisation and pruning erode alignment before they erode general benchmark scores, so compressed checkpoints answer harmful prompts more often. Robustness to rounding is the common assumption being refuted; the incoherence and calibration-set claims describe effects not reliably observe... |
pretraining_data_pipelines_0198 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Rephrased and synthetic pretraining text: paraphrasing web data with LLMs, textbook-style generation, and their benefits and failure modes | A pipeline generates synthetic pretraining documents with a model whose training data included popular benchmarks. Name the contamination risk this creates that string-matching decontamination of the seed corpus does not address. | {
"A": "The generator reproduces memorised benchmark items when prompted on related topics",
"B": "The generator copies verbatim spans from the seed documents it is shown",
"C": "The generator emits near-duplicate documents that survive MinHash deduplication",
"D": "The generator inherits toxic language present... | A | Decontaminating the seeds cannot remove benchmark content that lives in the generator's weights and is emitted during generation, so synthetic corpora need their own decontamination pass. Verbatim seed copying is caught by matching against seeds, near-duplicate collapse is a diversity issue, and toxicity is a separate ... |
safety_alignment_practice_0007 | Safety and Alignment Practice | safety_alignment_practice | mixed | Refusal training: harmful-request datasets, borderline cases and measuring over-refusal against under-refusal | Andriushchenko and Flammarion showed that a simple grammatical rewrite of a harmful request substantially raises compliance in several aligned chat models. Name the rewrite. | {
"A": "Reformulating the request in the past tense",
"B": "Appending a suffix optimised by greedy coordinate gradient",
"C": "Encoding the request in base64 before submitting it",
"D": "Splitting the request across several conversational turns"
} | A | Their result is that asking 'how did people make X?' rather than 'how do I make X?' evades refusal training, showing that refusal generalises poorly across tense. Base64 encoding, multi-turn splitting and GCG suffixes are all genuine jailbreak techniques but are not the past-tense finding being asked about. |
reasoning_models_rlvr_0096 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Test-time compute scaling: longer thinking, parallel sampling with majority vote or verifier selection, and their cost curves | In self-consistency (majority voting over k sampled chains of thought), accuracy typically rises quickly and then plateaus well before k reaches hundreds. What property of the sampled answer distribution causes this plateau? | {
"A": "The empirical mode converges to the model's most likely answer, so extra samples stop changing the winner",
"B": "The sampling temperature drifts upward with k, so later samples are drawn from a flatter distribution",
"C": "The chains grow longer with each sample, so later ones exceed the context window a... | A | Majority voting estimates the mode of the model's answer distribution; once that estimate is stable, more samples cannot help, so the ceiling is the mode's correctness rate. Temperature does not drift with k, truncation is unrelated to sample count, and answer-form splitting is a normalization issue that affects the vo... |
scaling_laws_compute_allocation_0140 | Scaling Laws and Compute Allocation | scaling_laws_compute_allocation | fundamentals | Data-constrained scaling: repeated epochs, the value of repeated data, and when repetition stops helping | A team has a fixed unique-token budget and far more compute than one Chinchilla-optimal pass would consume, so they train for many epochs. In the fitted repeated-data scaling framework, what is the mechanism that makes the loss they reach fall short of the Chinchilla prediction for that total processed-token count? | {
"A": "The excess-compute term in the law saturates once parameter count exceeds the data-constrained optimum, so capacity added beyond that point returns essentially nothing on held-out loss for the corpus.",
"B": "Allocation under a data constraint shifts compute toward parameters rather than passes, so the over... | D | The fitted data-constrained laws index loss by effective unique tokens, where the value of a repeated token decays with each pass until further epochs contribute nothing, so achieved loss lags the naive Chinchilla curve computed on raw processed tokens. The claim about an excess-capacity term saturating describes the p... |
interpretability_in_practice_0023 | Interpretability in Practice | interpretability_in_practice | modern | The residual stream and the linear representation hypothesis as the framing for most current techniques | In a sparse autoencoder trained on residual stream activations, what role do the columns of the decoder matrix play under the linear representation hypothesis? | {
"A": "They are the candidate feature directions in the residual stream that features are added along",
"B": "They are the detectors that decide whether each feature is active on a given input",
"C": "They are the sparsity penalties applied separately to each feature coefficient",
"D": "They are the residual e... | A | Reconstruction is a sparse non-negative combination of decoder columns, so each column is a claimed feature direction in activation space. Detection is the encoder's role; penalties are scalars in the loss; the residual error is not a matrix column. |
pretraining_data_pipelines_0058 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Heuristic quality filters (Gopher and C4 style rules) versus model-based quality classifiers trained on curated positives | The Gopher quality rules reject any document that does not contain at least two of the words 'the', 'be', 'to', 'of', 'and', 'that', 'have', 'with'. Which extraction-stage failure is this stopword test designed to catch? | {
"A": "Documents whose retained text is a table, a tag-soup fragment or a bare word list rather than connected sentences of running English prose",
"B": "Documents whose repeated navigation footer survived line-level cleanup and now accounts for a large share of the remaining token count",
"C": "Documents whose ... | A | A handful of English function words is the cheapest signal that the extracted string is grammatical running text at all, so the rule fires on tables, markup residue, keyword lists and other languages. Truncation after the lead paragraph leaves fluent prose that passes the stopword test easily and is instead an extracto... |
scaling_laws_compute_allocation_0238 | Scaling Laws and Compute Allocation | scaling_laws_compute_allocation | fundamentals | μP (Maximal Update Parametrisation) and muTransfer: width-based parameterisations that let hyperparameters found on small models transfer to large ones, at a conceptual level | Under μP, how is the learning rate for the input embedding layer scaled as width increases, compared with hidden weight matrices? | {
"A": "It is scaled down faster than hidden matrices, as one over width squared",
"B": "It is scaled down identically to hidden matrices, as one over width",
"C": "It is scaled up in width, whereas hidden matrices are held constant",
"D": "It is held constant in width, whereas hidden matrices are scaled down"
... | D | Embedding weights have fan-in equal to vocabulary size, which does not grow with width, so μP keeps their Adam learning rate width-independent while hidden matrices get a 1/width factor. Equal scaling, a 1/width-squared rule, and an upward-scaled embedding rate all misdescribe μP's per-layer table. |
synthetic_data_and_distillation_0133 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Filtering synthetic data: model judges, verifiers, consistency checks and deduplication of generated samples | In the classic STaR loop, when the model fails to reach the correct answer for a problem, the rationalisation step is applied before the sample can be added to the training set. Describe what that step does. | {
"A": "It reweights the failed sample's loss in proportion to its answer confidence",
"B": "It resamples the problem at a higher temperature until an answer matches",
"C": "It gives the model the gold answer as a hint and keeps the rationale it then produces",
"D": "It replaces the model's rationale with a rat... | C | STaR's rationalisation conditions on the correct answer as a hint, then keeps the generated reasoning (with the hint stripped) so hard problems are not lost. Higher-temperature resampling, teacher rewriting and loss reweighting are other plausible recovery strategies but not what rationalisation denotes. |
transformer_architecture_0074 | Decoder-Only Transformer Architecture | transformer_architecture | fundamentals | Positional encodings: learned absolute, sinusoidal, ALiBi and rotary embeddings, and how RoPE encodes relative position | Position Interpolation (linear RoPE scaling) extends a model's context by transforming the position indices in what way before the rotation matrices are applied? | {
"A": "Multiplying the RoPE base frequency by the extension factor raised to d/(d-2), leaving indices unchanged",
"B": "Scaling only the low-frequency rotary dimensions while leaving the high-frequency dimensions at their original wavelengths",
"C": "Dividing each position index by the extension factor so long s... | C | Linear Position Interpolation rescales indices m to m/s, compressing a longer sequence into the same range of rotation angles seen in pretraining, then briefly fine-tunes. Raising the base frequency by the extension factor to the d/(d-2) power is NTK-aware scaling, which explicitly leaves indices alone; scaling only lo... |
mixture_of_experts_0270 | Mixture-of-Experts Architectures | mixture_of_experts | modern | Upcycling dense checkpoints into MoE and other initialisation strategies for experts | In MoE upcycling, some recipes add small random perturbations to the duplicated expert weights instead of copying the dense MLP exactly. What property of the converted model is given up by that perturbation? | {
"A": "The saving from storing one shared copy of the expert weights plus small per-expert deltas, since identical tensors can no longer be deduplicated on disk and in device memory",
"B": "The ability to carry the dense run's optimiser moments across into the MoE run, since the moments no longer match the perturb... | C | Copying experts verbatim makes the top-k MoE compute the same function as the dense model at step zero, so the run resumes from the dense loss; perturbing the copies trades that exact equivalence for faster expert diversification. Router initialisation is unrelated to whether experts are perturbed, since the gate is a ... |
evaluation_methodology_0044 | Evaluation Methodology | evaluation_methodology | mixed | Scoring choices in harnesses: few-shot format, log-likelihood versus generative scoring, answer extraction, and perplexity comparability across tokenisers and its relation to downstream ability | State why a lower validation perplexity on a model's own pretraining distribution does not reliably predict which of two models will win on an instruction-following benchmark. | {
"A": "Perplexity is dominated by high-frequency tokens whose modelling is unrelated to the rare decisions a benchmark tests",
"B": "Perplexity is computed with teacher forcing, which prevents the accumulation of error across autoregressive steps",
"C": "Each model's held-out set reflects its own data mixture, s... | C | In-distribution perplexity is only comparable when the distribution is shared; each model's own validation split makes the numbers measure different targets. Teacher forcing, token frequency skew and outlier sensitivity are genuine properties of perplexity but do not explain a cross-model comparison being ill-posed. |
hardware_numerics_kernels_0262 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Compilers and graph optimisation (torch.compile, XLA) and their trade-offs against hand-written kernels | In torch.compile, mode="max-autotune" raises compile time relative to the default mode, which simply dispatches matmuls to cuBLAS. What is the extra work this mode performs? | {
"A": "It defers tuning until runtime, timing each Triton config on the first several calls before locking in a choice",
"B": "It reorders the operator schedule using profiled memory traces so that peak activation memory during backward drops",
"C": "It captures the compiled region into a CUDA graph so that repl... | D | max-autotune's defining cost is a compile-time benchmarking search: for each matmul shape it actually runs library kernels (cuBLAS/CUTLASS) against generated Triton templates and records the winner in the autotune cache. Capturing the region into a CUDA graph is what mode="reduce-overhead" does, not max-autotune; defer... |
optimisers_training_stability_0191 | Optimisers and Training Stability | optimisers_training_stability | fundamentals | Loss spikes: causes such as attention logit growth and bad batches, and remedies including QK-normalisation (the home for this mechanism), z-loss and checkpoint rewinding | A team using bf16 without z-loss finds their final-layer logits drifting to magnitudes of several hundred late in training, though the loss stays flat. What downstream numerical failure is this drift most likely to cause first? | {
"A": "Rounding of the second-moment buffer for the unembedding weights to zero",
"B": "Loss of the sign bit on negative logits during the bf16 to fp32 cast operation",
"C": "Underflow of the small softmax probabilities to exactly zero, giving infinite log-probabilities",
"D": "Overflow of the fp32 accumulator... | C | With max-subtraction the large exponentials are safe, but tail probabilities become so small that they underflow to zero, so their logs are -inf and gradients or evaluation metrics break. The fp32 accumulator after max-subtraction does not overflow, second moments are unaffected by logit scale, and casting bf16 to fp32... |
scaling_laws_compute_allocation_0197 | Scaling Laws and Compute Allocation | scaling_laws_compute_allocation | fundamentals | Scaling fits for hyperparameters: how optimal learning rate and batch size shift with model and data size (the critical-batch-size mechanism itself is covered under Optimisers) | In muTransfer, the hyperparameters tuned on a small proxy model are transferred to the large target model along the width axis. Which hyperparameter is explicitly NOT transferred in this way and must still be chosen for the large run? | {
"A": "The initialisation variance of the hidden weight matrices",
"B": "The peak learning rate of the optimiser",
"C": "The total training duration and its token budget",
"D": "The multiplier applied to the output logits"
} | C | muP prescribes width-scaling rules for init variance, LR and the output multiplier so those transfer; training length (and data budget) is not a width-parametrisation quantity and must be set separately, typically by compute-optimal scaling laws. The other three are precisely the quantities muP rescales and transfers. |
mixture_of_experts_0110 | Mixture-of-Experts Architectures | mixture_of_experts | modern | Fine-grained and shared experts: many small experts versus few large ones and the role of always-on shared experts | In the DeepSeekMoE ablation that removes the always-on shared expert while holding activated parameters per token constant by routing to one additional expert, what happens to validation loss? | {
"A": "Loss degrades markedly, so the extra routed expert fails to compensate for the lost shared capacity",
"B": "Loss is essentially unchanged after the router re-balances, since activated capacity per token is identical",
"C": "Loss degrades early in training but converges to the same value once expert specia... | A | The paper reports a clear degradation in validation loss when the shared expert is dropped at matched activated parameters, which is its evidence that unconditional common-knowledge capacity contributes beyond raw parameter count. The claim of no change after router re-balancing treats the shared expert as pure capacit... |
tokenisation_and_vocabularies_0317 | Tokenisation and Vocabularies | tokenisation_and_vocabularies | fundamentals | Alternatives to subword tokenisation: byte-level and character-level models, patching and dynamic tokenisation, and their costs | Multi-byte prediction is added to byte-level models such as BLT as a training and inference aid. State the mechanism. | {
"A": "Several output heads predict the next few bytes at once, letting more bytes be emitted per forward pass.",
"B": "Several byte embeddings are averaged into one input position, halving the sequence the model sees.",
"C": "Several candidate byte continuations are scored and the highest-probability patch is c... | A | Multi-token (here multi-byte) prediction attaches extra heads predicting positions t+1, t+2, ..., which densifies the training signal and supports self-speculative decoding, easing the serial-step problem of byte models. Input-side averaging is a downsampling scheme, beam-style patch commitment is not multi-byte predic... |
optimisers_training_stability_0056 | Optimisers and Training Stability | optimisers_training_stability | fundamentals | Learning-rate schedules: warmup, cosine decay, warmup-stable-decay and constant-with-cooldown, and their trade-offs | Two runs share a budget and peak rate, one on cosine to near zero and one on constant LR with no cooldown at all. What is the expected ordering of their final training losses? | {
"A": "They end at essentially the same loss, since cumulative learning rate is nearly equal",
"B": "The constant-LR run ends lower on training loss but higher on held-out loss",
"C": "The constant-LR run ends lower, having spent more steps at a productive rate",
"D": "The cosine run ends lower, because the an... | D | Without any decay, the constant-LR run remains at a high noise floor and plateaus above the annealed cosine run. Extra high-LR steps do not compensate for the missing anneal; the cumulative-LR argument alone ignores the annealing term; and the train/held-out split described is not the observed effect. |
long_context_and_retrieval_0162 | Long Context and Retrieval | long_context_and_retrieval | modern | Evaluating long context: needle-in-a-haystack, multi-needle and reasoning-over-context tests, and the lost-in-the-middle effect | In RULER's variable-tracking task the context contains chains such as X1 = 12345, X2 = X1, X3 = X2, and the model must report every variable bound to a given value. Beyond locating the relevant text spans, what ability does this task specifically require? | {
"A": "Following the assignments hop by hop, where each next position can only be found after the previous link is resolved",
"B": "Discriminating the queried value from numerically similar decoy values inserted at controlled distances throughout the filler text, which is what sets the difficulty level of the task... | A | The chained bindings make the task multi-hop: the position of the next assignment is unknown until the current alias is resolved, so retrieval alone is insufficient. Aggregating order-independent evidence describes multi-needle retrieval, where all needles can be found in parallel; numerically similar decoys and prefix... |
reasoning_models_rlvr_0192 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Overthinking and length control: length penalties, budget forcing and adjustable reasoning effort | In the s1 paper's budget forcing technique, what does the decoder do when the model tries to emit an end-of-thinking token before the allotted thinking budget is used up? | {
"A": "It injects the original question again as a reminder and lets the model restart its solution",
"B": "It suppresses the end-of-thinking token and appends the string \"Wait\" to prompt further reasoning",
"C": "It resamples the whole reasoning trace at a higher temperature until a longer trace appears",
"... | B | s1 suppresses the end-of-thinking delimiter and appends "Wait", which typically makes the model re-check its work. Resampling at higher temperature, backtracking to uncertain tokens, and re-injecting the prompt are plausible-sounding alternative decoding hacks but not what budget forcing does. |
transformer_architecture_0319 | Decoder-Only Transformer Architecture | transformer_architecture | fundamentals | Encoder-decoder and encoder-only models compared with decoder-only, and where each remains in use | State the reason decoder-only models scaled to hundreds of billions of parameters on plain web text while encoder-decoder denoising models largely did not follow to that scale. | {
"A": "The causal objective needs no tokenizer, letting raw byte streams be consumed directly at web scale",
"B": "The causal objective yields lower peak activation memory per layer in backpropagation, so larger batches fit on one device without recomputation",
"C": "The causal objective avoids the quadratic att... | D | Next-token prediction extracts a training signal at every position of any unlabelled document and requires no decisions about masking ratio, span length or sentinel format, so scaling on heterogeneous web corpora is uniform and cheap to configure, whereas denoising objectives supervise only the corrupted fraction and d... |
agents_and_tool_use_0299 | Agents and Tool Use | agents_and_tool_use | modern | Failure modes: looping, hallucinated tool results, cascading errors and premature termination | Sampling several candidate actions at each step and having the agent vote among them or verify them before execution raises token cost, but the failure mode it directly reduces is which one? | {
"A": "Premature termination inherited from single-turn instruction-tuning priors that reward emitting a final answer early",
"B": "Context overflow from long tool observations accumulating across many agent turns",
"C": "Cascading error from committing to one wrong action whose output conditions later steps",
... | C | Choosing among sampled candidates filters a bad action before it executes, so the erroneous observation never enters the context to corrupt subsequent reasoning, which is exactly the single-step-mistake propagation problem. Irreversible side effects persist because the voted-on action is still executed for real, and re... |
reasoning_models_rlvr_0151 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Reward hacking under verifiable rewards, as distinct from learned-reward-model hacking: test-case exploitation, answer-format gaming and reference-answer leakage | In RLVR on competition math, some problems ask for a proof but the reward only checks a final claimed statement. Name the resulting failure of the reward signal with respect to proof tasks. | {
"A": "Group advantages vanish on proofs, since claimed statements are almost never string-identical",
"B": "Symbolic normalizers reject proofs, since natural-language steps cannot be canonicalized at all",
"C": "Assertion without justification is fully rewarded, since the checkable target ignores the argument",... | C | Outcome checking on a claimed statement gives full credit for simply asserting it, which is why proof tasks are hard to put under RLVR. Length is not mechanically penalized, advantages do not automatically vanish, and normalizer behaviour is not the reason assertion succeeds. |
open_model_landscape_0131 | Open Model Landscape Since 2024 | open_model_landscape | modern | Reading technical reports: what is disclosed about data, compute and post-training, common omissions, and how to compare self-reported benchmark numbers across releases | A report gives its pretraining mixture only as broad category percentages and releases no corpus. Why does this make an independent contamination check for a newly published benchmark impossible? | {
"A": "Because measuring overlap requires searching the training text itself, which is not available.",
"B": "Because contamination is defined only against a benchmark's private test split, held by its authors.",
"C": "Because n-gram matching requires the benchmark to be tokenized with the model's own tokenizer.... | A | Overlap detection is a string- or n-gram-level search against the actual training documents, so without the corpus the check cannot be run at all. Contamination checks routinely use public test items, matching does not require the model's tokenizer, and loss curves are not the standard detector. |
pretraining_data_pipelines_0289 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Toxicity, personal-data and copyright filtering at the corpus stage and the trade-offs against coverage (memorisation, extraction attacks and unlearning are covered under Safety) | In the Dolma pipeline, detected email addresses, phone numbers and IP addresses inside otherwise-retained documents are handled how? | {
"A": "They are replaced in place with fixed placeholder strings marking the type",
"B": "They are hashed so that duplicate occurrences can still be counted",
"C": "They are left intact but the document is downweighted in the mixture",
"D": "They are shifted to random values so that surface statistics are pres... | A | Dolma masks a small number of detected PII spans with typed placeholder tokens and drops documents that exceed a threshold of detections. Hashing, downweighting and value randomisation are not what the released pipeline does. |
open_model_landscape_0235 | Open Model Landscape Since 2024 | open_model_landscape | modern | Multilingual and region-specific open models: the training decisions behind them (tokeniser design, language mixture, translated post-training data) and their trade-offs against general models | A team fine-tunes a Vietnamese-adapted derivative of a Llama base model and wants to release it. Under the Llama community licences, what obligation applies to the derivative's naming? | {
"A": "The derivative must be released under the identical licence text with no additional terms",
"B": "The derivative's name must include the base model family's name at the start",
"C": "The derivative must publish the fine-tuning dataset alongside the released weights",
"D": "The derivative may not be rele... | B | Llama's community licence requires derivative model names to begin with 'Llama' and to carry the 'Built with Llama' attribution. Commercial use is permitted below the licence's user threshold without a separate agreement. The licence is not a strict copyleft requiring identical redistribution terms, and it imposes no d... |
distributed_training_systems_0001 | Distributed Training Systems | distributed_training_systems | fundamentals | Data parallelism and gradient all-reduce: ring and tree algorithms, bandwidth cost and overlap with backward computation | Ring all-reduce is implemented as two phases over the ring. Name the two phases, in order. | {
"A": "All-gather of raw chunks, then a reduce-scatter pass",
"B": "Gather to a root rank, then scatter of averaged chunks",
"C": "Broadcast of the full buffer, then a local reduction pass",
"D": "Reduce-scatter, then all-gather of the reduced chunks"
} | D | Ring all-reduce first reduce-scatters so each rank owns one fully reduced chunk, then all-gathers those chunks. Reversing the order would communicate unreduced data and is not what NCCL does; gather-to-root then scatter is the parameter-server pattern; broadcast-then-local-reduce does not produce a global sum. |
pretraining_data_pipelines_0157 | Pretraining Data Pipelines | pretraining_data_pipelines | modern | Data mixture design: domain weights, upsampling code and math, and empirical or learned methods for setting proportions | In 'Scaling Data-Constrained Language Models', which extra data source do the authors mix into the pretraining set to fill a natural-language token deficit, reporting little degradation on non-code benchmarks? | {
"A": "Model-generated paraphrases of the C4 corpus, used at up to half the training mixture",
"B": "Sentence-aligned parallel translation corpora, used at up to half the training mixture",
"C": "LaTeX sources of arXiv papers from The Pile, used at up to half the training mixture",
"D": "Python source code dra... | D | The paper's data-augmentation experiments add Python code from The Stack, showing that up to roughly half the tokens can be code without hurting non-code task performance, which effectively multiplies the available token budget. Model-generated paraphrases are a later synthetic-rephrasing line of work, not this paper's... |
optimisers_training_stability_0003 | Optimisers and Training Stability | optimisers_training_stability | fundamentals | Adam and AdamW mechanics: moment estimates, bias correction, epsilon, decoupled weight decay (the home for weight-decay questions) and the role of beta2 | Beta2 controls the decay of Adam's second-moment estimate. Increasing beta2 from 0.95 to 0.999 changes the optimiser's response to a single anomalously large gradient in what way? | {
"A": "The denominator rises sharply for one step and then relaxes, briefly shrinking that parameter's update",
"B": "The denominator absorbs the spike only slightly but keeps it in the average for far more steps",
"C": "The denominator is unaffected, since outliers enter through the first moment instead",
"D"... | B | A larger beta2 gives each new squared gradient weight 1-beta2 = 0.001, so a spike barely moves the running average but persists in it over a horizon of roughly 1/(1-beta2) steps. The sharp-then-relax description fits small beta2; outliers do enter the second moment; epsilon saturation is a misreading of the floor's rol... |
synthetic_data_and_distillation_0194 | Synthetic Data and Distillation | synthetic_data_and_distillation | modern | Synthetic data for low-resource domains and languages, and translation-based data generation | Translating a benchmark such as MMLU into many languages and evaluating on the result is known to inflate apparent cross-lingual ability. Name the property of translated test items, relative to natively authored ones, that drives this inflation. | {
"A": "Translated items retain the English cultural, geographic and entity context the model has already absorbed",
"B": "Translated items lose the distractor calibration of the originals, leaving one option obviously implausible",
"C": "Translated items exhibit simplified syntax and a narrowed vocabulary that m... | A | Translation preserves the source culture's entities, units, laws and framing, so a model can answer from knowledge acquired in English pretraining without any locale-specific knowledge, which is exactly why natively authored multilingual suites score lower. Simplification and length inflation are genuine translationese... |
mixture_of_experts_0127 | Mixture-of-Experts Architectures | mixture_of_experts | modern | Fine-grained and shared experts: many small experts versus few large ones and the role of always-on shared experts | Given a fixed activated-parameter budget, what happens to a fine-grained MoE's total parameter count relative to a coarse-grained one when both use the same number of devices and the same per-expert placement, but the fine-grained model keeps the same total expert width? | {
"A": "It grows, because each split expert retains the original expert's full width",
"B": "It stays the same, because segmentation redistributes width without adding weights",
"C": "It shrinks, because narrower experts have fewer rows in each projection matrix",
"D": "It grows, because the router must store o... | B | Fine-grained segmentation partitions the same total intermediate width across more experts, so total expert parameters are conserved (aside from a negligibly larger router). Splitting does not duplicate full-width experts, does not remove weights, and the router stores one vector per expert, not per token bucket. |
distributed_training_systems_0098 | Distributed Training Systems | distributed_training_systems | fundamentals | Pipeline parallelism: micro-batching, bubble overhead, 1F1B and interleaved schedules, and what zero-bubble schedules change at a conceptual level | In an interleaved 1F1B pipeline schedule where each device owns v non-contiguous virtual stages, by what factor is the bubble fraction reduced relative to non-interleaved 1F1B with the same p and m? | {
"A": "By a factor of v², since warm-up and drain each shrink by v independently, and the two reductions compound at the ends.",
"B": "By a factor of √v, since the shorter per-stage computation is partly offset by the extra boundary crossings every micro-batch incurs.",
"C": "By a factor of p/v, since the count ... | D | Splitting each device's layers into v chunks makes every pipeline step 1/v as long, so the p−1 warm-up and drain steps cost 1/v of the original idle time, giving a bubble fraction of (p−1)/(vm) instead of (p−1)/m. The v² claim double-counts a single reduction that applies once to the combined warm-up/drain time; the √v... |
preference_optimisation_rlhf_0059 | Preference Optimisation and RLHF | preference_optimisation_rlhf | mixed | Reward model training: Bradley–Terry objective, architecture on top of the policy, and reward-model evaluation | Suppose a reward model is trained only on comparisons between completions sampled from an early SFT checkpoint, then used to score a heavily PPO-optimised policy. What property of the reward model is being violated? | {
"A": "Its calibration to the annotators' pairwise agreement rate",
"B": "Its transitivity over triples of candidate responses",
"C": "Its invariance to the ordering of responses within a pair",
"D": "Its accuracy on inputs drawn from the policy's current distribution"
} | D | Reward models generalise reliably only near the distribution their training comparisons came from, so an off-policy RM degrades as the policy drifts — the reason for iterative RM refresh with fresh on-policy comparisons. Order invariance holds by construction for a scalar scorer, transitivity is automatic for scalar re... |
agents_and_tool_use_0054 | Agents and Tool Use | agents_and_tool_use | modern | Agent loops: ReAct-style interleaving of reasoning and actions, scratchpads and termination conditions | A ReAct agent emits an action line that fails to parse against the harness's expected format. Name the conventional harness response that keeps the episode alive. | {
"A": "Resampling the same step at a higher temperature without recording the malformed text",
"B": "Charging the step to the iteration budget and returning an empty observation string",
"C": "Truncating the scratchpad back to the last well-formed turn and resuming generation",
"D": "Appending a parse-error ob... | D | The standard convention treats a malformed action exactly like a failed tool call: the error text becomes the next observation in the scratchpad, so the model sees its own mistake in context and re-emits a well-formed action. Truncating back to the last well-formed turn discards the very evidence the model needs and is... |
transformer_architecture_0239 | Decoder-Only Transformer Architecture | transformer_architecture | fundamentals | Bias terms and dropout in large-scale pretraining: why both were largely removed and what took over their roles | Beyond weight decay, which data-side practice is most often credited with taking over dropout's regularising role in large-scale LLM pretraining? | {
"A": "Packing documents to fill the full context window",
"B": "Upsampling high-quality sources such as books and code",
"C": "Deduplicating the corpus so examples are rarely repeated",
"D": "Shuffling documents so each batch mixes many sources"
} | C | Deduplication directly limits memorisation by ensuring near-duplicate text is not seen many times, which is the failure mode dropout addressed. Shuffling, packing and quality upsampling are all standard pipeline steps but target batch composition, throughput and data quality rather than repetition-driven overfitting. |
preference_optimisation_rlhf_0037 | Preference Optimisation and RLHF | preference_optimisation_rlhf | mixed | Reward model training: Bradley–Terry objective, architecture on top of the policy, and reward-model evaluation | RewardBench evaluates reward models by measuring one particular quantity across curated prompt sets. What does it measure? | {
"A": "The accuracy with which the model assigns a higher score to the chosen response",
"B": "The KL divergence between the reward model's and the policy's output distributions",
"C": "The downstream win rate of policies trained by PPO against the reward model",
"D": "The correlation between reward scores and... | A | RewardBench is a pairwise-accuracy benchmark: a preference pair is scored correct if the chosen completion receives the higher reward. Absolute Likert correlation, KL to a policy and downstream PPO win rates are other evaluation ideas, but not what this benchmark reports. |
distributed_training_systems_0315 | Distributed Training Systems | distributed_training_systems | fundamentals | Fault tolerance at scale: hardware failures, straggler detection, asynchronous checkpointing, restart determinism and elastic restarts | Why does increasing the number of accelerators in a job, holding per-device reliability constant, shorten the job's mean time between interruptions? | {
"A": "Synchronous collectives make any single device's fault fatal, since no replica can absorb the loss",
"B": "Independent per-device failure hazards add, so the aggregate failure rate scales with device count",
"C": "Larger jobs span more racks and switches, so network faults grow superlinearly with device c... | B | For independent components the union failure rate is the sum of the individual rates, so time to first failure falls roughly inversely with the number of devices; that arithmetic is the reason. Gang-scheduled synchronous training does make one fault fatal to the step, but that explains the blast radius rather than the ... |
mixture_of_experts_0037 | Mixture-of-Experts Architectures | mixture_of_experts | modern | Router design: softmax gating, expert-choice routing, sigmoid scoring and normalisation of routing weights | In Shazeer et al.'s noisy top-k gating, tunable Gaussian noise is added to the router logits before selection. State the purpose of that noise. | {
"A": "To keep router logits small enough to remain stable in low-precision arithmetic",
"B": "To decorrelate expert outputs so their combination has lower variance",
"C": "To help load balancing by letting near-threshold experts sometimes be selected",
"D": "To smooth the top-k boundary so gradients pass thro... | C | The noise makes routing stochastic near the decision boundary, encouraging exploration so experts other than the current favourites get traffic, which supports the load-balancing losses. It does not create gradients for unselected experts, is not a precision-control device, and is not aimed at output variance. |
hardware_numerics_kernels_0166 | Hardware, Numerics and Kernels | hardware_numerics_kernels | mixed | Sub-8-bit number formats (MXFP4, NVFP4, INT4) as formats: block scaling, dynamic range and hardware support | NVFP4 applies a second, per-tensor FP32 scale on top of its per-16-element E4M3 block scales. State the purpose of that second level of scaling. | {
"A": "It stores the per-channel zero points required for asymmetric quantisation of activations",
"B": "It compensates for missing subnormal encodings in E2M1, restoring resolution near zero for small weights",
"C": "It normalises the tensor so all block scales land inside the representable E4M3 range",
"D": ... | C | E4M3 block scales have a bounded exponent range, so a single FP32 factor rescales the whole tensor to bring every block's required scale into that encodable range. E2M1 does include a subnormal representation, so nothing about missing subnormals is being compensated; FP32 accumulation is a tensor-core datapath property... |
agents_and_tool_use_0110 | Agents and Tool Use | agents_and_tool_use | modern | Computer-use and browser agents: screenshot or DOM observation, action spaces and grounding errors | For desktop computer-use agents, what is the main advantage cited for a pixel-level action space (click at x,y, type, scroll, key press) over an element-id action space? | {
"A": "It lets the environment check each action against the current widget hierarchy before executing it, so clicks on stale or offscreen targets are rejected rather than misapplied",
"B": "It generalises to any application, including canvas-rendered and native UIs that expose no queryable element tree",
"C": "... | B | Pixel actions need nothing from the application beyond rendered pixels, so they cover games, PDF and video viewers, remote desktops and native toolkits with no accessibility tree, which is the usual justification. Claiming coordinates remove grounding error inverts the trade-off: mapping a described target to x,y is ex... |
reasoning_models_rlvr_0155 | Reasoning Models and RLVR | reasoning_models_rlvr | modern | Reward hacking under verifiable rewards, as distinct from learned-reward-model hacking: test-case exploitation, answer-format gaming and reference-answer leakage | In multi-turn RLVR environments where the model can query the grader for feedback between attempts, name the exploit that arises when the number of queries is unbounded. | {
"A": "Deliberately failing early turns so the group's reward variance stays high throughout training",
"B": "Emitting the same patch every turn so the grader caches the verdict and returns it faster",
"C": "Splitting the reasoning across turns so each turn stays under the per-turn token budget",
"D": "Brute-f... | D | Unlimited oracle queries turn the grader into an information source, letting the model infer hidden expectations rather than derive a solution, which is why query budgets are imposed. Deliberate failure, caching and token budgeting do not extract reward-relevant information from the grader. |
interpretability_in_practice_0263 | Interpretability in Practice | interpretability_in_practice | modern | Attention pattern analysis: induction heads, positional and copying heads, and what attention maps do not show | In the IOI (indirect object identification) circuit studied in GPT-2 small, name mover heads copy the correct name to the final position. What role do the S-inhibition heads play upstream of them? | {
"A": "They detect the sentence boundary so name movers restrict attention to the second clause",
"B": "They negate the name mover output when the subject appears twice in the prompt",
"C": "They write a signal that suppresses name mover attention to the duplicated subject token",
"D": "They write the duplicat... | C | S-inhibition heads modify the name movers' queries so they attend away from the repeated subject and toward the indirect object. Copying the subject forward is the opposite of their effect; boundary detection is not their described function; negating the mover output describes negative name movers instead. |
inference_and_serving_0161 | Inference and Serving Systems | inference_and_serving | mixed | Post-training weight quantisation methods (GPTQ, AWQ, SmoothQuant style) and calibration-data effects | AWQ chooses per-channel scaling factors for weight quantisation using activation magnitude statistics rather than weight magnitudes. What claim about which weights matter motivates this choice? | {
"A": "Weight channels with the largest absolute values dominate the layer's quantisation error, so per-channel weight maxima should set the scales.",
"B": "Weight channels multiplied by large-magnitude activation outliers dominate the layer's quantisation error, so their relative error must be shrunk first.",
"... | B | AWQ's salience criterion is activation-side: channels hit by outlier activations contribute most to output error, so scaling them up before rounding reduces their relative quantisation error. Selecting by largest absolute weight value is precisely the naive baseline AWQ measures against and beats. Hessian-diagonal sens... |
inference_and_serving_0038 | Inference and Serving Systems | inference_and_serving | mixed | KV cache mechanics and memory sizing, paged/blocked allocation, and prefix caching for shared-prompt reuse across requests | A chat service replays a long shared system prompt on every request. When prefix caching hits, which latency metric improves directly? | {
"A": "Inter-token latency, because decode attends over fewer cached key-value entries",
"B": "Queueing delay, because cache hits let the scheduler admit requests out of order",
"C": "Total decode time, because sampled tokens are reused along with the prefix keys",
"D": "Time to first token, because the cached... | D | A prefix hit removes prefill work for the matched tokens, which is precisely the dominant term in time to first token. Decode still attends over the full context, so per-token latency is unchanged; scheduling order is not altered by a hit; generated tokens are never reused from the cache. |
preference_optimisation_rlhf_0006 | Preference Optimisation and RLHF | preference_optimisation_rlhf | mixed | Preference data collection: pairwise comparisons, rating scales, annotator agreement and AI-generated preferences | Anthropic's helpful-and-harmless preference collection used an interface where the crowdworker chose between two model responses at each turn of an ongoing conversation, and the chosen response continued the dialogue. What was the stated purpose of collecting preferences inside multi-turn dialogues rather than on singl... | {
"A": "To reduce annotation cost by amortising the reading of shared conversational context over several successive comparisons, so each additional label is cheaper",
"B": "To let annotators revisit and revise their earlier turn-level judgements once the full conversation had been completed and rated",
"C": "To ... | D | Because the selected response is appended and the dialogue continues, later comparisons are drawn from the same conversational distribution the model generates and encounters in deployment, which was the stated motivation. Cost amortisation is at best an incidental side effect, not the rationale given; the interface of... |
scaling_laws_compute_allocation_0278 | Scaling Laws and Compute Allocation | scaling_laws_compute_allocation | fundamentals | Scaling laws for specialised settings: MoE sparsity, vocabulary size and distillation as additional axes, and how numeric precision enters loss fits (the formats themselves are under Hardware) | Beyond the parameter cost of embeddings, what practical training cost grows with vocabulary size and is often cited as a reason not to push vocabularies to the compute-optimal value? | {
"A": "Tokenising a fixed corpus yields fewer tokens overall, so a fixed compute budget covers proportionally more epochs of the data",
"B": "The attention key-value cache grows per layer because more distinct token types must be kept distinguishable in memory",
"C": "The output softmax logits tensor grows with ... | C | The logits tensor is batch × sequence × vocabulary, so its materialisation and the cross-entropy over it dominate activation memory once V is large, which is why chunked or fused loss kernels are used and why very large vocabularies are avoided in practice. Undertrained rare-token embeddings are a genuine phenomenon bu... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.