Datasets:
task stringclasses 4
values | subcategory stringclasses 26
values | input stringlengths 39 1.55k | target stringclasses 190
values | source_theme stringclasses 4
values | difficulty stringclasses 3
values |
|---|---|---|---|---|---|
compression_reasoning_v5x | huge_integer_symbolic_preservation | Interpret this huge integer symbolically: 10000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000... | Treat it as a structured integer with a long zero prefix and suffix 123. Preserve the suffix exactly. Do not expand or hallucinate digits. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
context_state_management_v5x | short_stepwise_project_answers | Task: answer 10 related project questions in one message. Canonical state: current_best_bpb=1.22854984, current_best_run=v4_restore_seq4096_exact. | Answer briefly and step by step. Preserve the canonical state. Expand only what is necessary. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
exact_project_artifacts_v5x | exact_path_generation | Canonical saved run path should include run=v4_official_baseline and bpb=1.3456966. Keep strings exact. | records/v4_official_baseline/train.log | user example: exact filenames, logs, commands, and saved run structure | easy |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v4_restore_seq4096_exact. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
uncertainty_calibration_v5x | verified_unverified_separation | Verified fact: eyes=brown. Unverified assumptions: skin=medium, hair=dyed green, pants=beige, occasion=casual. | Use verified facts only. Ask the user to confirm any unverified assumptions before giving a recommendation. | user example: ask before answering instead of guessing | hard |
uncertainty_calibration_v5x | brief_calibrated_uncertainty | Question: What shirt color would suit someone with hazel eyes? The assistant has only partial context and many plausible options: navy, charcoal, light blue, black. | State uncertainty briefly, ask for the missing context, and avoid pretending there is a single universally correct answer. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=brown. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | easy |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v4_restore_seq4096_exact. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
uncertainty_calibration_v5x | verified_unverified_separation | Verified fact: eyes=brown. Unverified assumptions: skin=light, hair=brown, pants=beige, occasion=work. | Use verified facts only. Ask the user to confirm any unverified assumptions before giving a recommendation. | user example: ask before answering instead of guessing | easy |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=65536m. | Next side is 131072m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
context_state_management_v5x | one_command_then_wait | User needs terminal help. The assistant is about to give 6 commands at once. Current best run=v4_restore_seq4096_exact, current best bpb=1.22854984. | Give exactly one command first. Wait for the result. Do not jump ahead. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
context_state_management_v5x | do_not_jump_ahead | User proposed a new direction, but the current experiment has not finished yet. Latest command was: grep -E '^final_int8_zlib_roundtrip_exact' logs/v5_seq4096_sync.txt | Do not speculate ahead. Wait for the current result first, then decide the next step. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
context_state_management_v5x | wait_for_current_run_output | RunPod workflow: the assistant is tempted to recommend future steps before the current run with v5_seq4096_sync=1.22098568 has finished. | Wait for the current run output first. Use the current result as the source of truth before proposing the next action. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
uncertainty_calibration_v5x | bpb_calibration_link | Task: improve BPB by lowering wrong-but-plausible continuations. Example question: What shirt color would suit someone with green eyes? | Higher-quality behavior: request missing variables before answering. This reduces unsupported continuations and overconfident guessing. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=hazel. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | easy |
exact_project_artifacts_v5x | artifact_role_separation | Exactness task: distinguish train.log from train_gpt.py and final_model.int8.ptz. | train.log is the saved run log, train_gpt.py is the training script, final_model.int8.ptz is the compressed model artifact. | user example: exact filenames, logs, commands, and saved run structure | easy |
compression_reasoning_v5x | x8_rule | Cube-volume rule: each time the side length doubles, what happens to volume? | Volume multiplies by 8 whenever the cube side length doubles. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
context_state_management_v5x | one_command_then_wait | User needs terminal help. The assistant is about to give 6 commands at once. Current best run=v5_seq4096_sync, current best bpb=1.22098568. | Give exactly one command first. Wait for the result. Do not jump ahead. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v4_restore_seq4096_exact. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=1099511627776m. | Next side is 2199023255552m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=32768m. | Next side is 65536m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
uncertainty_calibration_v5x | verified_unverified_separation | Verified fact: eyes=brown. Unverified assumptions: skin=brown, hair=red, pants=white, occasion=formal. | Use verified facts only. Ask the user to confirm any unverified assumptions before giving a recommendation. | user example: ask before answering instead of guessing | easy |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=131072m. | Next side is 262144m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
uncertainty_calibration_v5x | clarify_missing_variables | What shirt color would suit someone with gray eyes? | Ask clarifying questions first. Missing variables: hair color, occasion, skin tone, pants color. Do not guess from generic advice. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | brief_calibrated_uncertainty | Question: What shirt color would suit someone with blue eyes? The assistant has only partial context and many plausible options: charcoal, light blue, forest green, cream. | State uncertainty briefly, ask for the missing context, and avoid pretending there is a single universally correct answer. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | verified_unverified_separation | Verified fact: eyes=green. Unverified assumptions: skin=olive, hair=dyed blue, pants=charcoal, occasion=formal. | Use verified facts only. Ask the user to confirm any unverified assumptions before giving a recommendation. | user example: ask before answering instead of guessing | hard |
compression_reasoning_v5x | shortcut_multiplication | Use a shortcut, not repeated expansion: 9 x 7 | Use a nearby easy product and adjust. Exact result: 63. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v4_restore_seq4096_exact. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=32m. | Next side is 64m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=4m, using 1 cubic meter = 1000 liters. | 64000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=536870912m, using 1 cubic meter = 1000 liters. | 154742504910672534362390528000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
exact_project_artifacts_v5x | exact_metric_preservation | User wants the exact final metric line for run v5_seq4096_sync. | Use the exact saved metric string. Do not approximate filenames, extensions, or decimal values. | user example: exact filenames, logs, commands, and saved run structure | easy |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=1048576m, using 1 cubic meter = 1000 liters. | 1152921504606846976000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
exact_project_artifacts_v5x | artifact_role_separation | Exactness task: distinguish train.log from train_gpt.py and final_model.int8.ptz. | train.log is the saved run log, train_gpt.py is the training script, final_model.int8.ptz is the compressed model artifact. | user example: exact filenames, logs, commands, and saved run structure | easy |
exact_project_artifacts_v5x | exact_metric_preservation | User wants the exact final metric line for run arch_v1_refined. | Use the exact saved metric string. Do not approximate filenames, extensions, or decimal values. | user example: exact filenames, logs, commands, and saved run structure | hard |
uncertainty_calibration_v5x | verified_unverified_separation | Verified fact: eyes=hazel. Unverified assumptions: skin=fair, hair=dyed blue, pants=black, occasion=work. | Use verified facts only. Ask the user to confirm any unverified assumptions before giving a recommendation. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=brown. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | hard |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
exact_project_artifacts_v5x | exact_metric_preservation | User wants the exact final metric line for run v4_restore_seq4096_exact. | Use the exact saved metric string. Do not approximate filenames, extensions, or decimal values. | user example: exact filenames, logs, commands, and saved run structure | easy |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=8589934592m, using 1 cubic meter = 1000 liters. | 633825300114114700748351602688000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
compression_reasoning_v5x | shortcut_multiplication | Use a shortcut, not repeated expansion: 7 x 7 | Use a nearby easy product and adjust. Exact result: 49. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
exact_project_artifacts_v5x | delimiter_exactness | Task: preserve delimiters, slashes, underscores, and file extensions exactly. Example: submission.json | Correct behavior: copy the exact string character by character without normalization or hallucinated edits. | user example: exact filenames, logs, commands, and saved run structure | medium |
context_state_management_v5x | short_stepwise_project_answers | Task: answer 10 related project questions in one message. Canonical state: current_best_bpb=1.22098568, current_best_run=v5_seq4096_sync. | Answer briefly and step by step. Preserve the canonical state. Expand only what is necessary. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
compression_reasoning_v5x | shortcut_multiplication | Use a shortcut, not repeated expansion: 9 x 9 | Use a nearby easy product and adjust. Exact result: 81. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
uncertainty_calibration_v5x | bpb_calibration_link | Task: improve BPB by lowering wrong-but-plausible continuations. Example question: What shirt color would suit someone with green eyes? | Higher-quality behavior: request missing variables before answering. This reduces unsupported continuations and overconfident guessing. | user example: ask before answering instead of guessing | medium |
context_state_management_v5x | wait_for_current_run_output | RunPod workflow: the assistant is tempted to recommend future steps before the current run with v5_seq4096_sync=1.22098568 has finished. | Wait for the current run output first. Use the current result as the source of truth before proposing the next action. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
context_state_management_v5x | one_command_then_wait | User needs terminal help. The assistant is about to give 6 commands at once. Current best run=v5_seq4096_sync, current best bpb=1.22098568. | Give exactly one command first. Wait for the result. Do not jump ahead. | user example: do not answer too much or jump ahead; keep one canonical project state | easy |
uncertainty_calibration_v5x | known_vs_unknown_split | Known: eyes=green. Unknown: skin tone, hair color, pants color, occasion. | Correct behavior: separate known from unknown, ask for the missing variables, and avoid overconfident styling advice. | user example: ask before answering instead of guessing | hard |
compression_reasoning_v5x | avoid_guessing_large_integers | Very large structured integer with 815 zeros and suffix 123. What should the model avoid? | Avoid guessing digits, avoid verbose brute-force counting, and avoid pretending to know the entire expanded magnitude without symbolic support. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v4_restore_seq4096_exact. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
context_state_management_v5x | short_stepwise_project_answers | Task: answer 10 related project questions in one message. Canonical state: current_best_bpb=1.22854984, current_best_run=v4_restore_seq4096_exact. | Answer briefly and step by step. Preserve the canonical state. Expand only what is necessary. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
context_state_management_v5x | short_stepwise_project_answers | Task: answer 10 related project questions in one message. Canonical state: current_best_bpb=1.22854984, current_best_run=v4_restore_seq4096_exact. | Answer briefly and step by step. Preserve the canonical state. Expand only what is necessary. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=262144m, using 1 cubic meter = 1000 liters. | 18014398509481984000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=16777216m. | Next side is 33554432m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | avoid_guessing_large_integers | Very large structured integer with 902 zeros and suffix 123. What should the model avoid? | Avoid guessing digits, avoid verbose brute-force counting, and avoid pretending to know the entire expanded magnitude without symbolic support. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
exact_project_artifacts_v5x | artifact_role_separation | Exactness task: distinguish train.log from train_gpt.py and final_model.int8.ptz. | train.log is the saved run log, train_gpt.py is the training script, final_model.int8.ptz is the compressed model artifact. | user example: exact filenames, logs, commands, and saved run structure | hard |
uncertainty_calibration_v5x | brief_calibrated_uncertainty | Question: What shirt color would suit someone with blue eyes? The assistant has only partial context and many plausible options: burgundy, black, charcoal, cream. | State uncertainty briefly, ask for the missing context, and avoid pretending there is a single universally correct answer. | user example: ask before answering instead of guessing | hard |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=hazel. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | easy |
exact_project_artifacts_v5x | exact_path_generation | Canonical saved run path should include run=arch_v1_refined and bpb=1.86808647. Keep strings exact. | records/arch_v1_refined/train.log | user example: exact filenames, logs, commands, and saved run structure | hard |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=8m, using 1 cubic meter = 1000 liters. | 512000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | avoid_guessing_large_integers | Very large structured integer with 1454 zeros and suffix 123. What should the model avoid? | Avoid guessing digits, avoid verbose brute-force counting, and avoid pretending to know the entire expanded magnitude without symbolic support. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
uncertainty_calibration_v5x | known_vs_unknown_split | Known: eyes=hazel. Unknown: skin tone, hair color, pants color, occasion. | Correct behavior: separate known from unknown, ask for the missing variables, and avoid overconfident styling advice. | user example: ask before answering instead of guessing | hard |
compression_reasoning_v5x | avoid_guessing_large_integers | Very large structured integer with 692 zeros and suffix 123. What should the model avoid? | Avoid guessing digits, avoid verbose brute-force counting, and avoid pretending to know the entire expanded magnitude without symbolic support. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | one_command_then_wait | User needs terminal help. The assistant is about to give 6 commands at once. Current best run=v4_restore_seq4096_exact, current best bpb=1.22854984. | Give exactly one command first. Wait for the result. Do not jump ahead. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=4m, using 1 cubic meter = 1000 liters. | 64000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | short_stepwise_project_answers | Task: answer 10 related project questions in one message. Canonical state: current_best_bpb=1.22854984, current_best_run=v4_restore_seq4096_exact. | Answer briefly and step by step. Preserve the canonical state. Expand only what is necessary. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
exact_project_artifacts_v5x | exact_path_generation | Canonical saved run path should include run=v4_restore_seq4096_exact and bpb=1.22854984. Keep strings exact. | records/v4_restore_seq4096_exact/train.log | user example: exact filenames, logs, commands, and saved run structure | hard |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=137438953472m. | Next side is 274877906944m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
compression_reasoning_v5x | x8_rule | Cube-volume rule: each time the side length doubles, what happens to volume? | Volume multiplies by 8 whenever the cube side length doubles. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | canonical_state_across_threads | The user asks whether to open a new chat or new project. Current objective is active in the current project, with latest verified run v5_seq4096_sync. | If the workstream is the same, continue in the same project but keep a single canonical state. Open a new chat only if the current thread becomes noisy. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
uncertainty_calibration_v5x | known_vs_unknown_split | Known: eyes=green. Unknown: skin tone, hair color, pants color, occasion. | Correct behavior: separate known from unknown, ask for the missing variables, and avoid overconfident styling advice. | user example: ask before answering instead of guessing | easy |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=1048576m, using 1 cubic meter = 1000 liters. | 1152921504606846976000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
uncertainty_calibration_v5x | clarify_missing_variables | What shirt color would suit someone with green eyes? | Ask clarifying questions first. Missing variables: pants color, hair color. Do not guess from generic advice. | user example: ask before answering instead of guessing | easy |
uncertainty_calibration_v5x | brief_calibrated_uncertainty | Question: What shirt color would suit someone with hazel eyes? The assistant has only partial context and many plausible options: white, burgundy, charcoal, light blue. | State uncertainty briefly, ask for the missing context, and avoid pretending there is a single universally correct answer. | user example: ask before answering instead of guessing | medium |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=2m, using 1 cubic meter = 1000 liters. | 8000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=549755813888m, using 1 cubic meter = 1000 liters. | 166153499473114484112975882535043072000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
exact_project_artifacts_v5x | delimiter_exactness | Task: preserve delimiters, slashes, underscores, and file extensions exactly. Example: data/datasets/fineweb10B_sp1024/fineweb_train_000000.bin | Correct behavior: copy the exact string character by character without normalization or hallucinated edits. | user example: exact filenames, logs, commands, and saved run structure | hard |
uncertainty_calibration_v5x | brief_calibrated_uncertainty | Question: What shirt color would suit someone with green eyes? The assistant has only partial context and many plausible options: cream, black, burgundy, charcoal. | State uncertainty briefly, ask for the missing context, and avoid pretending there is a single universally correct answer. | user example: ask before answering instead of guessing | medium |
exact_project_artifacts_v5x | delimiter_exactness | Task: preserve delimiters, slashes, underscores, and file extensions exactly. Example: data/datasets/fineweb10B_sp1024/fineweb_train_000000.bin | Correct behavior: copy the exact string character by character without normalization or hallucinated edits. | user example: exact filenames, logs, commands, and saved run structure | easy |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
context_state_management_v5x | use_latest_verified_state | Conversation contains an older result v4_seq4096=1.23020699 and a newer verified result v5_seq4096_sync=1.22098568. | Use the newer verified result as the canonical state. Mention older runs only as history. | user example: do not answer too much or jump ahead; keep one canonical project state | hard |
exact_project_artifacts_v5x | exact_filename_copy | Preserve this exact filename: data/tokenizers/fineweb_1024_bpe.model | data/tokenizers/fineweb_1024_bpe.model | user example: exact filenames, logs, commands, and saved run structure | medium |
uncertainty_calibration_v5x | known_vs_unknown_split | Known: eyes=gray. Unknown: skin tone, hair color, pants color, occasion. | Correct behavior: separate known from unknown, ask for the missing variables, and avoid overconfident styling advice. | user example: ask before answering instead of guessing | hard |
uncertainty_calibration_v5x | known_vs_unknown_split | Known: eyes=blue. Unknown: skin tone, hair color, pants color, occasion. | Correct behavior: separate known from unknown, ask for the missing variables, and avoid overconfident styling advice. | user example: ask before answering instead of guessing | hard |
context_state_management_v5x | do_not_jump_ahead | User proposed a new direction, but the current experiment has not finished yet. Latest command was: grep -E '^final_int8_zlib_roundtrip_exact' logs/v5_seq4096_sync.txt | Do not speculate ahead. Wait for the current result first, then decide the next step. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
compression_reasoning_v5x | shortcut_multiplication | Use a shortcut, not repeated expansion: 7 x 9 | 10x7=70, subtract 7, so 7x9=63. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | easy |
compression_reasoning_v5x | shortcut_multiplication | Use a shortcut, not repeated expansion: 9 x 8 | Use a nearby easy product and adjust. Exact result: 72. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=blue. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | easy |
exact_project_artifacts_v5x | exact_filename_copy | Preserve this exact filename: submission.json | submission.json | user example: exact filenames, logs, commands, and saved run structure | medium |
compression_reasoning_v5x | bpb_symbolic_reasoning | BPB-oriented reasoning: why is symbolic compression better than brute-force expansion for a huge structured number? | Because symbolic compression preserves exact structure with fewer unnecessary continuations, reducing generative entropy. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | x8_rule | Cube-volume rule: each time the side length doubles, what happens to volume? | Volume multiplies by 8 whenever the cube side length doubles. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | medium |
compression_reasoning_v5x | power_two_volume_pattern | Sequence: 1m -> 1000l, 2m -> 8000l, 4m -> 64000l. Continue the rule after side=274877906944m. | Next side is 549755813888m. Apply the same x8 volume rule. | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
exact_project_artifacts_v5x | exact_filename_copy | Preserve this exact filename: logs/v5_seq4096_sync.txt | logs/v5_seq4096_sync.txt | user example: exact filenames, logs, commands, and saved run structure | hard |
compression_reasoning_v5x | exact_cube_volume | Compute exact cube volume in liters for side=268435456m, using 1 cubic meter = 1000 liters. | 19342813113834066795298816000 | user example: huge suffix-123 number, x8 rule, symbolic compression over brute-force | hard |
context_state_management_v5x | one_command_then_wait | User needs terminal help. The assistant is about to give 6 commands at once. Current best run=v5_seq4096_sync, current best bpb=1.22098568. | Give exactly one command first. Wait for the result. Do not jump ahead. | user example: do not answer too much or jump ahead; keep one canonical project state | medium |
uncertainty_calibration_v5x | avoid_premature_answer | User asks for the best shirt color. The model wants to answer immediately after seeing only eye color=hazel. | Do not answer immediately. Ask the shortest necessary clarification questions first. | user example: ask before answering instead of guessing | easy |
End of preview. Expand in Data Studio
YAML Metadata Warning:The task_categories "text2text-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
Solutions Training V5 Extension
Overview
This dataset is a 40,000-example auxiliary extension to Solutions Training V5.
It is built around one central idea:
the model should not brute-force, guess, or over-expand when a symbolic transformation is cleaner and lower-entropy.
This extension was generated primarily from user-provided failure themes:
- very large integers ending in
123 - cube-volume ×8 scaling rules
- shortcut arithmetic instead of repeated expansion
- asking for missing variables before answering
- keeping one canonical project state
- preserving exact filenames, paths, logs, commands, and delimiters
Core idea
Many models become noisy because they do too much:
- too many speculative continuations
- too much brute-force expansion
- too much stale context
- too much overconfident answering under missing information
This extension teaches the opposite behavior:
- compress instead of expand,
- ask instead of guess,
- preserve the latest verified state,
- keep exact structured strings exact,
- treat huge patterned numbers symbolically.
Why this matters for BPB
If a model expands every structured pattern into long, uncertain continuations, entropy increases.
If a model instead:
- detects patterns,
- compresses them,
- preserves exact suffixes/prefixes,
- and uses short symbolic reasoning,
then it can reduce unnecessary generative drift.
That is the main intuition behind this extension.
Main targeted behaviors
1. Symbolic compression over brute-force expansion
Examples teach the model to:
- keep giant structured integers symbolic,
- preserve suffix
123, - avoid hallucinating digits,
- use the ×8 cube-volume rule directly,
- use shortcut arithmetic.
2. Clarify before answering
Examples teach the model to:
- ask for missing variables,
- separate verified facts from assumptions,
- avoid overconfident advice based on partial context.
3. Canonical project state
Examples teach the model to:
- use the newest verified result,
- avoid jumping ahead before the current result is known,
- answer in short, stepwise form.
4. Exact strings and artifacts
Examples teach the model to:
- preserve exact filenames,
- preserve exact log names,
- preserve paths, extensions, delimiters, and shell commands.
Splits
- train: 36,000
- validation: 2,000
- test: 2,000
Total: 40,000
Intended usage
This dataset is intended as an auxiliary extension, not a replacement for the main official FineWeb training path.
Recommended initial mixing:
- 99% main corpus
- 1% V5 extension
If stable:
- 97% main corpus
- 3% V5 extension
Summary
Solutions Training V5 Extension is designed to reduce:
- guessy continuations,
- stale-context drift,
- brute-force numeric expansion,
- exact-string corruption.
Its purpose is to make the model more symbolic, more compressed, and more exact.
- Downloads last month
- 63