id stringlengths 10 10 | category stringclasses 3
values | task stringclasses 18
values | instruction stringclasses 16
values | input stringlengths 23 138 | target stringlengths 1 183 | rationale stringclasses 26
values | difficulty stringclasses 2
values | language stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|
pgcce_0195 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | If the best run uses TRAIN_SEQ_LEN=4096, should TRAIN_SEQ_LEN=1024 be treated as identical? | No. The sequence length is a different configuration. | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0185 | exact_discrete_reasoning | exact_arithmetic | Compute exact arithmetic and prefer short exact decompositions over vague reasoning. | Compute 14 x 9 using a short exact method. | 126. One exact shortcut is 10 x 14 = 140, then 140 - 14 = 126. | The model should preserve exactness in small discrete calculations. | easy | en |
pgcce_0063 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: My training setup is broken in case 13. What should I do? | Ask for the exact traceback, current environment version, and the last command that failed. Fix one verified issue at a time. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0024 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know a question mentions a cube, but the user has not specified whether the shape is a cube or an approximation to a sphere. | Verified: a geometric container is discussed. Unverified: exact shape model. Ask whether the estimate should use cube volume or a rounded approximation. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0167 | exact_discrete_reasoning | exact_copy | Copy exact filenames, delimiters, and command tokens with zero changes. | Copy the exact command token without modification: MAX_WALLCLOCK_SECONDS=0 | MAX_WALLCLOCK_SECONDS=0 | BPB improves when the model preserves exact discrete forms instead of paraphrasing them. | easy | en |
pgcce_0067 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: My computer feels slow in scenario 17. What should I upgrade? | Ask for current specs, workload, bottleneck symptoms, and budget. Upgrades should not be guessed from one vague complaint. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0108 | context_state_management | canonical_state_priority | Resolve long-context ambiguity by preferring the current canonical state. | Canonical state says a filename typo caused the earlier copy failure. | Use the verified filename from the sorted log list. | The model should compress project memory into a stable source of truth. | medium | en |
pgcce_0069 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: What is the best option for me in scenario 19? | State that the request is underspecified, ask for the key constraints, and avoid a generic one-size-fits-all answer. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0107 | context_state_management | canonical_state_priority | Resolve long-context ambiguity by preferring the current canonical state. | Canonical state says README is outdated but logs are current. | Trust the verified logs first, then update the README. | The model should compress project memory into a stable source of truth. | medium | en |
pgcce_0140 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 30: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0161 | exact_discrete_reasoning | exact_copy | Copy exact filenames, delimiters, and command tokens with zero changes. | Copy the exact file path without modification: logs/v4_seq4096.txt | logs/v4_seq4096.txt | BPB improves when the model preserves exact discrete forms instead of paraphrasing them. | easy | en |
pgcce_0158 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 131072 m and volume 2251799813685248000 liters. What is the volume at side length 262144 m? | 18014398509481984000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | medium | en |
pgcce_0199 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | If a question asks for the exact current state, should older chat fragments override the newest verified run? | No. Use the newest verified run as the source of truth. | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0142 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 2 m and volume 8000 liters. What is the volume at side length 4 m? | 64000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | easy | en |
pgcce_0032 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: Which skincare routine is right for me? | I should not guess yet. Please tell me your skin type, sensitivity, current routine, and main concern. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0128 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 18: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0058 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: My training setup is broken in case 8. What should I do? | Ask for the exact traceback, current environment version, and the last command that failed. Fix one verified issue at a time. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0113 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 3: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0157 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 65536 m and volume 281474976710656000 liters. What is the volume at side length 131072 m? | 2251799813685248000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | medium | en |
pgcce_0075 | context_state_management | use_latest_verified_state | Answer from the latest verified project state, not from older fragments. | The project had multiple pods. Only one was stopped, not deleted. The user asks whether work can continue. | State the current verified pod state first: stopped pods preserve workspace data, deleted pods do not. | Long context should be compressed into one canonical state before reasoning continues. | medium | en |
pgcce_0102 | context_state_management | canonical_state_priority | Resolve long-context ambiguity by preferring the current canonical state. | Canonical state says best_run=arch_v1_refined is historical V3, while V4 official baseline is 1.3457 and V4 seq4096 is 1.2302. | Treat V4 seq4096 as current best and V3 as historical reference. | The model should compress project memory into a stable source of truth. | medium | en |
pgcce_0134 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 24: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0152 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 2048 m and volume 8589934592000 liters. What is the volume at side length 4096 m? | 68719476736000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | medium | en |
pgcce_0036 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: Which keyboard switch is best? | I should not guess yet. Please tell me your noise tolerance, typing vs gaming, force preference, and layout. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0175 | exact_discrete_reasoning | count_and_preserve_structure | Count exactly or preserve the exact order and delimiters. | How many slash-separated sides are written here: 1m/1m/1m/1m/1m/1m/ | 6 | The model should treat counts, separators, and order as first-class constraints. | easy | en |
pgcce_0149 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 256 m and volume 16777216000 liters. What is the volume at side length 512 m? | 134217728000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | easy | en |
pgcce_0138 | context_state_management | summarize_then_next_action | Maintain project state and reduce long-context drift. | Scenario 28: The user asks for the next action after a partial success in a long debugging session. | Summarize the verified current state in one sentence, then give exactly one next action. | This teaches state compression before planning. | medium | en |
pgcce_0133 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 23: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0085 | context_state_management | one_step_at_a_time | Answer with the next necessary action only, then wait for confirmation. | The user asks how to restore a run. | Step 1: extract the exact config from the winning log. Do not rerun with guessed defaults. | Short stepwise execution reduces compounding context errors. | easy | en |
pgcce_0112 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 2: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0089 | context_state_management | one_step_at_a_time | Answer with the next necessary action only, then wait for confirmation. | The user asks how to save the best run before deleting a pod. | Step 1: copy log, script, and final artifact into a dedicated records folder. | Short stepwise execution reduces compounding context errors. | easy | en |
pgcce_0164 | exact_discrete_reasoning | exact_copy | Copy exact filenames, delimiters, and command tokens with zero changes. | Copy the exact file path without modification: ./data/datasets/fineweb10B_sp1024 | ./data/datasets/fineweb10B_sp1024 | BPB improves when the model preserves exact discrete forms instead of paraphrasing them. | easy | en |
pgcce_0189 | exact_discrete_reasoning | exact_arithmetic | Compute exact arithmetic and prefer short exact decompositions over vague reasoning. | Compute 64 x 8 exactly. | 512. | The model should preserve exactness in small discrete calculations. | easy | en |
pgcce_0110 | context_state_management | canonical_state_priority | Resolve long-context ambiguity by preferring the current canonical state. | Canonical state says one file contains the current best log. | Use that file and stop searching older fragments. | The model should compress project memory into a stable source of truth. | medium | en |
pgcce_0001 | uncertainty_calibration | clarify_before_recommendation | Answer with calibrated uncertainty. Ask for the missing variables before recommending. | User: Which shirt color will suit blue eyes best? | Clarify first: What is your skin tone, hair color, trouser color, and occasion? Eye color alone is not enough to choose the best shirt. | The question is underspecified. Ask about skin tone, hair color, trouser color, and occasion before recommending a color or style. | easy | en |
pgcce_0030 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know a dataset upload failed because of YAML metadata. | Verified: the upload failed. Unverified: every cause. Check the YAML header syntax first before changing other files. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0121 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 11: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0173 | exact_discrete_reasoning | count_and_preserve_structure | Count exactly or preserve the exact order and delimiters. | Return the list in the same order: warmup, train, validate, save | warmup, train, validate, save | The model should treat counts, separators, and order as first-class constraints. | easy | en |
pgcce_0016 | uncertainty_calibration | clarify_before_technical_advice | Ask short, high-value clarification questions before giving technical recommendations. | User: Should I use Docker here? | I need more context before answering. Please specify deployment target, team workflow, OS, and reproducibility needs. Do not choose a tool from incomplete requirements. | The model should ask focused clarification questions instead of guessing a generic answer. | easy | en |
pgcce_0126 | context_state_management | summarize_then_next_action | Maintain project state and reduce long-context drift. | Scenario 16: The user asks for the next action after a partial success in a long debugging session. | Summarize the verified current state in one sentence, then give exactly one next action. | This teaches state compression before planning. | medium | en |
pgcce_0040 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: Which monitor should I buy? | I should not guess yet. Please tell me your budget, gaming or work, size, and refresh rate requirement. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0010 | uncertainty_calibration | clarify_before_recommendation | Answer with calibrated uncertainty. Ask for the missing variables before recommending. | User: Which watch strap color suits me best? | Clarify first: What is your skin tone, wardrobe palette, formal or casual use, and metal color? Accessory advice should be calibrated, not guessed. | The question is underspecified. Ask about skin tone, wardrobe palette, formal or casual use, and metal color before recommending a color or style. | easy | en |
pgcce_0088 | context_state_management | one_step_at_a_time | Answer with the next necessary action only, then wait for confirmation. | The user asks how to add files to Hugging Face from a phone. | Step 1: fix README YAML. Step 2: upload the corrected files. Stop before discussing training. | Short stepwise execution reduces compounding context errors. | easy | en |
pgcce_0181 | exact_discrete_reasoning | exact_arithmetic | Compute exact arithmetic and prefer short exact decompositions over vague reasoning. | Compute 7 x 9 using a short exact method. | 63. One exact shortcut is 10 x 7 = 70, then 70 - 7 = 63. | The model should preserve exactness in small discrete calculations. | easy | en |
pgcce_0200 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | If final_model.int8.ptz is overwritten by later runs, can it be assumed to match any older log automatically? | No. It must be matched or regenerated for the specific run. | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0038 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: Which backpack should I buy? | I should not guess yet. Please tell me your capacity need, travel or daily use, budget, and weather exposure. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0074 | context_state_management | use_latest_verified_state | Answer from the latest verified project state, not from older fragments. | A conversation contains many ideas, but only one verified state is current: best run, best BPB, and artifact status. | Summarize the canonical state first, then answer. Do not continue from scattered fragments. | Long context should be compressed into one canonical state before reasoning continues. | medium | en |
pgcce_0124 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 14: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0015 | uncertainty_calibration | clarify_before_technical_advice | Ask short, high-value clarification questions before giving technical recommendations. | User: What programming language should I learn? | I need more context before answering. Please specify goals, current level, time available, and job target. Do not choose a tool from incomplete requirements. | The model should ask focused clarification questions instead of guessing a generic answer. | easy | en |
pgcce_0023 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know a repository contains logs/v4_seq4096.txt, but you have not checked whether final_model.int8.ptz matches that run. | Verified: the log file exists. Unverified: whether the current model artifact matches that log. Re-run or verify before claiming consistency. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0026 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know a command failed, but you do not yet know the exact traceback. | Verified: a failure happened. Unverified: the root cause. Ask for the exact error message before suggesting multiple fixes. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0050 | uncertainty_calibration | clarify_or_defer | If important fields are missing, ask first or explicitly defer a confident answer. | User: Which dataset split should I trust? | Before answering, clarify split definition, latest verification, data leakage status, and evaluation purpose. If those details are missing, state the uncertainty instead of guessing. | This teaches the model to prefer calibrated uncertainty over confident guessing. | medium | en |
pgcce_0066 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: I have blue eyes and need style advice case 16. What should I wear? | Ask for skin tone, hair color, trouser color, and occasion before answering. Do not infer the full outfit from eye color alone. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0065 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: Which color should I choose for project 15? | Ask about the existing palette, lighting, material, and target mood before recommending a color. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0087 | context_state_management | one_step_at_a_time | Answer with the next necessary action only, then wait for confirmation. | The user asks how to prepare V4 from V3. | Start with the official baseline-compatible stack, then carry over only the proven best run data. | Short stepwise execution reduces compounding context errors. | easy | en |
pgcce_0096 | context_state_management | short_high_signal_response | Prefer short, high-signal answers that protect state integrity. | The user asks what changed between two runs. | Report only the changed variables and the resulting BPB difference. | Concise answers reduce long-context drift and execution mistakes. | easy | en |
pgcce_0176 | exact_discrete_reasoning | count_and_preserve_structure | Count exactly or preserve the exact order and delimiters. | How many explicit variables appear here: current_best_bpb, current_best_run, submission_status | 3 | The model should treat counts, separators, and order as first-class constraints. | easy | en |
pgcce_0165 | exact_discrete_reasoning | exact_copy | Copy exact filenames, delimiters, and command tokens with zero changes. | Copy the exact file path without modification: final_model.int8.ptz | final_model.int8.ptz | BPB improves when the model preserves exact discrete forms instead of paraphrasing them. | easy | en |
pgcce_0132 | context_state_management | summarize_then_next_action | Maintain project state and reduce long-context drift. | Scenario 22: The user asks for the next action after a partial success in a long debugging session. | Summarize the verified current state in one sentence, then give exactly one next action. | This teaches state compression before planning. | medium | en |
pgcce_0039 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: How should I organize my repo? | I should not guess yet. Please tell me your project size, team size, release cadence, and compliance needs. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0035 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: What course should I take? | I should not guess yet. Please tell me your goal, current level, time available, and domain. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0002 | uncertainty_calibration | clarify_before_recommendation | Answer with calibrated uncertainty. Ask for the missing variables before recommending. | User: Which lipstick shade will look best on me? | Clarify first: What is your skin undertone, hair color, eye color, and day or evening use? A makeup recommendation should not be guessed from one trait. | The question is underspecified. Ask about skin undertone, hair color, eye color, and day or evening use before recommending a color or style. | easy | en |
pgcce_0003 | uncertainty_calibration | clarify_before_recommendation | Answer with calibrated uncertainty. Ask for the missing variables before recommending. | User: What haircut should I get? | Clarify first: What is your face shape, hair texture, hair length, and maintenance preference? A haircut recommendation depends on structure and preference. | The question is underspecified. Ask about face shape, hair texture, hair length, and maintenance preference before recommending a color or style. | easy | en |
pgcce_0042 | uncertainty_calibration | clarify_or_defer | If important fields are missing, ask first or explicitly defer a confident answer. | User: Is this answer definitely true? | Before answering, clarify source quality, date, whether the claim was verified here, and uncertainty level. If those details are missing, state the uncertainty instead of guessing. | This teaches the model to prefer calibrated uncertainty over confident guessing. | medium | en |
pgcce_0061 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: I have blue eyes and need style advice case 11. What should I wear? | Ask for skin tone, hair color, trouser color, and occasion before answering. Do not infer the full outfit from eye color alone. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0123 | context_state_management | summarize_then_next_action | Maintain project state and reduce long-context drift. | Scenario 13: The user asks for the next action after a partial success in a long debugging session. | Summarize the verified current state in one sentence, then give exactly one next action. | This teaches state compression before planning. | medium | en |
pgcce_0054 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: What is the best option for me in scenario 4? | State that the request is underspecified, ask for the key constraints, and avoid a generic one-size-fits-all answer. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0136 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 26: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0076 | context_state_management | use_latest_verified_state | Answer from the latest verified project state, not from older fragments. | The user asks 10 questions in one message about a training issue. | Answer in short ordered steps. Resolve the first blocker before proposing later steps. | Long context should be compressed into one canonical state before reasoning continues. | medium | en |
pgcce_0193 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | Artifact size is 12715304 bytes and total submission size is 12762990 bytes. Is it under 16000000 bytes? | Yes, it is under 16000000 bytes. | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0029 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know the phone calculator overflowed on large powers of two. | Verified: the calculator lost scale. Unverified: the final numeric result unless computed directly. Recompute from the recurrence instead of guessing. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0021 | uncertainty_calibration | separate_verified_from_unverified | Separate confirmed facts from missing or unverified information. | You know only that the user has blue eyes. You do not know skin tone, hair color, or outfit. | Verified: blue eyes. Unverified: skin tone, hair color, trouser color, occasion. Ask for the unverified fields before recommending. | The model should allocate probability mass to confirmed facts and avoid overcommitting to unverified claims. | medium | en |
pgcce_0119 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 9: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0187 | exact_discrete_reasoning | exact_arithmetic | Compute exact arithmetic and prefer short exact decompositions over vague reasoning. | Compute 16 x 8 exactly. | 128. | The model should preserve exactness in small discrete calculations. | easy | en |
pgcce_0191 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | Current best V4 exact BPB is 1.23020699 and historical V3 BPB is 1.86808647. Which is the current best? | 1.23020699 is the current best V4 BPB. | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0144 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 8 m and volume 512000 liters. What is the volume at side length 16 m? | 4096000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | easy | en |
pgcce_0062 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: My computer feels slow in scenario 12. What should I upgrade? | Ask for current specs, workload, bottleneck symptoms, and budget. Upgrades should not be guessed from one vague complaint. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0156 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 32768 m and volume 35184372088832000 liters. What is the volume at side length 65536 m? | 281474976710656000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | medium | en |
pgcce_0106 | context_state_management | canonical_state_priority | Resolve long-context ambiguity by preferring the current canonical state. | Canonical state says the pod was stopped, not deleted. | Continue from the existing pod instead of rebuilding unnecessarily. | The model should compress project memory into a stable source of truth. | medium | en |
pgcce_0198 | exact_discrete_reasoning | exact_project_state_reasoning | Use exact comparisons and file-state reasoning. | If a YAML field contains a colon in the value, what exact syntax avoids parser failure? | Wrap the value in quotes, for example: pretty_name: "Name: With Colon". | The model must preserve exact project facts and not blur distinct files or metrics. | medium | en |
pgcce_0115 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 5: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0077 | context_state_management | use_latest_verified_state | Answer from the latest verified project state, not from older fragments. | The assistant previously mixed two different run names. | Before giving advice, restate the current best verified run name and match it to the correct log file. | Long context should be compressed into one canonical state before reasoning continues. | medium | en |
pgcce_0177 | exact_discrete_reasoning | count_and_preserve_structure | Count exactly or preserve the exact order and delimiters. | Repeat exactly: 2m/2m/2m/2m/2m/2m | 2m/2m/2m/2m/2m/2m | The model should treat counts, separators, and order as first-class constraints. | easy | en |
pgcce_0122 | context_state_management | limit_future_branching | Maintain project state and reduce long-context drift. | Scenario 12: The assistant is tempted to give five future commands at once. | Give only the next command and wait for the outcome before proposing more. | This reduces long-chain execution errors. | medium | en |
pgcce_0146 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 32 m and volume 32768000 liters. What is the volume at side length 64 m? | 262144000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | easy | en |
pgcce_0114 | context_state_management | summarize_then_next_action | Maintain project state and reduce long-context drift. | Scenario 4: The user asks for the next action after a partial success in a long debugging session. | Summarize the verified current state in one sentence, then give exactly one next action. | This teaches state compression before planning. | medium | en |
pgcce_0049 | uncertainty_calibration | clarify_or_defer | If important fields are missing, ask first or explicitly defer a confident answer. | User: Should I quantize this tensor? | Before answering, clarify artifact limit, roundtrip loss impact, tensor sensitivity, and fallback option. If those details are missing, state the uncertainty instead of guessing. | This teaches the model to prefer calibrated uncertainty over confident guessing. | medium | en |
pgcce_0098 | context_state_management | short_high_signal_response | Prefer short, high-signal answers that protect state integrity. | The user asks how to resume work after stopping a pod. | Answer short: yes, if the pod was stopped, workspace data remains. | Concise answers reduce long-context drift and execution mistakes. | easy | en |
pgcce_0060 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: Which color should I choose for project 10? | Ask about the existing palette, lighting, material, and target mood before recommending a color. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0082 | context_state_management | one_step_at_a_time | Answer with the next necessary action only, then wait for confirmation. | The user needs to upload a dataset and README fails YAML parsing. | Step 1: fix the YAML header. Do not propose downstream training changes before the upload succeeds. | Short stepwise execution reduces compounding context errors. | easy | en |
pgcce_0139 | context_state_management | latest_metric_priority | Maintain project state and reduce long-context drift. | Scenario 29: There are multiple historical BPB values in memory, but only one newest verified result. | Use the newest verified BPB only and mark older values as historical. | This teaches the model to resist anchoring on stale metrics. | medium | en |
pgcce_0100 | context_state_management | short_high_signal_response | Prefer short, high-signal answers that protect state integrity. | The user asks how to avoid hallucination in recommendations. | State one rule: ask for missing high-impact variables first. | Concise answers reduce long-context drift and execution mistakes. | easy | en |
pgcce_0037 | uncertainty_calibration | calibrated_uncertainty_response | Respond briefly, admit missing information, and ask for the highest-value missing fields. | User: What diet should I follow? | I should not guess yet. Please tell me your goal, restrictions, activity level, and food preference. After that I can give a calibrated recommendation. | The right behavior is calibrated uncertainty followed by targeted clarification. | easy | en |
pgcce_0056 | uncertainty_calibration | underspecified_request_handling | Detect when a request is underspecified and ask the minimum set of high-impact questions. | User: I have blue eyes and need style advice case 6. What should I wear? | Ask for skin tone, hair color, trouser color, and occasion before answering. Do not infer the full outfit from eye color alone. | The model should not reward itself for sounding decisive when the problem statement is incomplete. | medium | en |
pgcce_0006 | uncertainty_calibration | clarify_before_recommendation | Answer with calibrated uncertainty. Ask for the missing variables before recommending. | User: Which sneakers should I wear with this outfit? | Clarify first: What is your trouser color, top color, occasion, and style preference? Outfit advice should request the missing variables first. | The question is underspecified. Ask about trouser color, top color, occasion, and style preference before recommending a color or style. | easy | en |
pgcce_0092 | context_state_management | short_high_signal_response | Prefer short, high-signal answers that protect state integrity. | The user is on a phone and cannot manage six code blocks at once. | Provide one command at a time, with one expected outcome. | Concise answers reduce long-context drift and execution mistakes. | easy | en |
pgcce_0078 | context_state_management | use_latest_verified_state | Answer from the latest verified project state, not from older fragments. | The user asks whether a dataset artifact and log belong to the same run. | Check the canonical run state first. Do not assume the current final_model.int8.ptz matches an older log. | Long context should be compressed into one canonical state before reasoning continues. | medium | en |
pgcce_0045 | uncertainty_calibration | clarify_or_defer | If important fields are missing, ask first or explicitly defer a confident answer. | User: What city has the best weather? | Before answering, clarify temperature preference, humidity tolerance, budget, and job location. If those details are missing, state the uncertainty instead of guessing. | This teaches the model to prefer calibrated uncertainty over confident guessing. | medium | en |
pgcce_0160 | exact_discrete_reasoning | cube_volume_x8 | Compute cube volume scaling exactly. Preserve the recurrence and the units. | A cube has side length 524288 m and volume 144115188075855872000 liters. What is the volume at side length 1048576 m? | 1152921504606846976000 liters. Doubling cube side length multiplies volume by 8. | The model must apply the exact x8 rule when side length doubles. | medium | en |
pgcce_0020 | uncertainty_calibration | clarify_before_technical_advice | Ask short, high-value clarification questions before giving technical recommendations. | User: What framework should I use? | I need more context before answering. Please specify project type, team size, latency needs, and deployment target. Do not choose a tool from incomplete requirements. | The model should ask focused clarification questions instead of guessing a generic answer. | easy | en |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.