Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
task_id
string
family
string
task_type
string
level
string
pass
int64
teacher_pass_rate
float64
input_ids
list
prompt_length
int64
length
int64
top_ids
list
top_logprobs
list
top_ids_privileged
list
top_logprobs_privileged
list
coverage
float64
math/gsm8k/train/287f9dd64ae1#0
math/gsm8k/train/287f9dd64ae1
math
gsm8k
default
1
1
[ 248045, 8678, 198, 2523, 11290, 1396, 3274, 539, 25773, 424, 430, 449, 24460, 3377, 13, 271, 16, 13, 52465, 25, 1528, 22898, 1092, 264, 4252, 466, 1786, 4087, 1902, 10752, 11, 321, 1204, 488, 1000, 3184, 264, 2577, 4087, 494, 264, 106...
253
587
[ [ 16, 332, 1206, 8160, 13962, 9764, 59, 40, 75620, 27775, 8046, 1596, 13682, 17, 1919, 71093, 550, 248068, 97490, 20 ], [ 13, 15, 17, 18261, 16, 8, 52465, 11, 21, 159029, 75620, 18, ...
[ [ -0.267578125, -2.017578125, -3.142578125, -3.517578125, -3.642578125, -6.515625, -6.890625, -10.515625, -11.015625, -11.015625, -11.140625, -11.390625, -11.390625, -11.515625, -11.828125, -11.828125, -12.078125, -12.078125, -12.328125, ...
null
null
0.998614
math/gsm8k/train/287f9dd64ae1#1
math/gsm8k/train/287f9dd64ae1
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
253
702
[[16,332,1206,8160,13962,9764,59,40,75620,27775,8046,1596,13682,17,1919,71093,550,248068,97490,20],[(...TRUNCATED)
[[-0.267578125,-2.017578125,-3.142578125,-3.517578125,-3.642578125,-6.515625,-6.890625,-10.515625,-1(...TRUNCATED)
null
null
0.999736
math/gsm8k/train/287f9dd64ae1#2
math/gsm8k/train/287f9dd64ae1
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
253
710
[[16,332,1206,8160,13962,9764,59,40,75620,27775,8046,1596,13682,17,1919,71093,550,248068,97490,20],[(...TRUNCATED)
[[-0.267578125,-2.017578125,-3.142578125,-3.517578125,-3.642578125,-6.515625,-6.890625,-10.515625,-1(...TRUNCATED)
null
null
0.999846
math/gsm8k/train/6da2feba679b#0
math/gsm8k/train/6da2feba679b
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
251
598
[[16,332,1206,13962,8160,59,9764,8046,75620,17,40,27775,20,13682,550,1919,71093,951,15,6527],[13,15,(...TRUNCATED)
[[-0.145263671875,-2.76953125,-3.26953125,-3.89453125,-4.39453125,-6.51953125,-9.1484375,-10.0234375(...TRUNCATED)
null
null
0.999238
math/gsm8k/train/6da2feba679b#1
math/gsm8k/train/6da2feba679b
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
251
604
[[16,332,1206,13962,8160,59,9764,8046,75620,17,40,27775,20,13682,550,1919,71093,951,15,6527],[13,15,(...TRUNCATED)
[[-0.145263671875,-2.76953125,-3.26953125,-3.89453125,-4.39453125,-6.51953125,-9.1484375,-10.0234375(...TRUNCATED)
null
null
0.999593
math/gsm8k/train/6da2feba679b#2
math/gsm8k/train/6da2feba679b
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
251
612
[[16,332,1206,13962,8160,59,9764,8046,75620,17,40,27775,20,13682,550,1919,71093,951,15,6527],[16,756(...TRUNCATED)
[[-0.145263671875,-2.76953125,-3.26953125,-3.89453125,-4.39453125,-6.51953125,-9.1484375,-10.0234375(...TRUNCATED)
null
null
0.998595
math/gsm8k/train/9b656e446524#0
math/gsm8k/train/9b656e446524
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
261
665
[[16,332,1206,13962,8160,59,9764,75620,8046,40,17,71093,13682,27775,1919,20,550,1596,248068,951],[13(...TRUNCATED)
[[-0.294189453125,-1.9189453125,-2.919921875,-3.294921875,-4.16796875,-6.66796875,-8.796875,-10.5468(...TRUNCATED)
null
null
0.999729
math/gsm8k/train/9b656e446524#1
math/gsm8k/train/9b656e446524
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
261
678
[[16,332,1206,13962,8160,59,9764,75620,8046,40,17,71093,13682,27775,1919,20,550,1596,248068,951],[13(...TRUNCATED)
[[-0.294189453125,-1.9189453125,-2.919921875,-3.294921875,-4.16796875,-6.66796875,-8.796875,-10.5468(...TRUNCATED)
null
null
0.999175
math/gsm8k/train/9b656e446524#2
math/gsm8k/train/9b656e446524
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
261
775
[[16,332,1206,13962,8160,59,9764,75620,8046,40,17,71093,13682,27775,1919,20,550,1596,248068,951],[13(...TRUNCATED)
[[-0.294189453125,-1.9189453125,-2.919921875,-3.294921875,-4.16796875,-6.66796875,-8.796875,-10.5468(...TRUNCATED)
null
null
0.999627
math/gsm8k/train/af072475e8ca#0
math/gsm8k/train/af072475e8ca
math
gsm8k
default
1
1
[248045,8678,198,2523,11290,1396,3274,539,25773,424,430,449,24460,3377,13,271,16,13,52465,25,1528,22(...TRUNCATED)
256
619
[[16,332,1206,8160,13962,59,9764,8046,75620,40,17,6527,20,97490,71093,550,27775,1919,248068,15],[13,(...TRUNCATED)
[[-0.267822265625,-2.017578125,-3.017578125,-3.392578125,-4.01953125,-7.01953125,-8.140625,-10.51562(...TRUNCATED)
null
null
0.996967
End of preview. Expand in Data Studio

Optimizer Distillation Data

Offline distillation data from Qwen/Qwen3.6-27B (teacher) for Qwen/Qwen3.5-0.8B (student). Every response solves its task as an optimization problem (objective, variables and constraints, search, verify, generalize), was graded correct by an automatic checker, and comes with the teacher's top-20 log-probabilities at every response token.

102,750 responses for 39,522 tasks, 63.3M scored response tokens, mean top-20 coverage 0.998; 9.8% of responses are pass-2 rescues.

Subsets

  • scores (default): one row per response, ready for training. input_ids holds the student's chat template (system prompt, task, an empty <think></think> block) followed by the response. Row t of top_ids / top_logprobs is the teacher's prediction for input_ids[prompt_length + t], normalized over the full vocabulary, so 1 - exp(top_logprobs).sum(-1) is the probability mass outside the top 20. coverage is that top-20 mass averaged over the row. top_ids_privileged / top_logprobs_privileged are the same scores under the privileged prompt (the task plus its gold answer), present only for pass == 2 rows.
  • responses: the same responses as text, with the task prompt, gold answer, checker score, and the teacher's thinking and response lengths.
Column Meaning
id <task id>#<sample index>
family, task_type, level Task family (math, puzzles, combinatorial), type and difficulty level
pass 1: the teacher solved the plain task; 2: rescued with the gold answer in the teacher prompt (see below)
teacher_pass_rate Share of the teacher's 3 plain-prompt samples for this task that were correct
input_ids, prompt_length, length Token ids for the student tokenizer; the response starts at prompt_length
top_ids, top_logprobs (response length, 20) teacher top-k token ids (int32) and log-probabilities (float16)
coverage Mean top-20 probability mass per response token

Counts by task

Family Task type Level Tasks Responses Pass-2 share
combinatorial knapsack large 2,089 4,113 18%
combinatorial knapsack medium 2,085 5,022 2%
combinatorial knapsack small 2,086 6,069 0%
combinatorial knapsack tiny 2,076 6,180 0%
combinatorial tsp medium 1,771 3,069 91%
combinatorial tsp small 2,076 4,487 20%
combinatorial tsp tiny 2,088 5,917 0%
math competition 1 915 2,718 0%
math competition 2 942 2,781 1%
math competition 3 935 2,710 1%
math competition 4 944 2,687 2%
math competition 5 913 2,381 4%
math gsm8k default 7,077 20,808 2%
puzzles countdown default 858 1,809 35%
puzzles countdown easy 856 2,207 3%
puzzles course_schedule default 859 2,373 0%
puzzles course_schedule easy 693 2,052 0%
puzzles jugs default 472 992 88%
puzzles knights_knaves default 856 2,506 0%
puzzles knights_knaves hard 857 2,484 0%
puzzles largest_island default 736 2,005 1%
puzzles maze default 794 1,456 51%
puzzles maze easy 855 1,812 8%
puzzles mini_sudoku default 864 2,487 0%
puzzles mini_sudoku easier 854 2,455 0%
puzzles mini_sudoku easy 853 2,441 0%
puzzles n_queens default 738 1,538 35%
puzzles rotten_oranges default 676 1,137 91%
puzzles shortest_path default 857 1,817 45%
puzzles shortest_path easy 847 2,237 2%

How it was made

  1. Tasks. GSM8K (openai/gsm8k) and competition math (qwedsacf/competition_math, levels 1–5), procedurally generated reasoning-gym puzzles, and procedurally generated knapsack and travelling-salesman instances with exact solvers. Puzzle and combinatorial difficulty levels were calibrated against the student.
  2. Generation. The teacher, served with vLLM in thinking mode (temperature 1.0, top-p 0.95, top-k 20, up to 12,288 new tokens), answered every task 3 times under the system prompt below. A response was kept when it finished, was graded correct (score ≥ 0.99), had a visible reasoning section before its final \boxed{}, and fit in 1,536 response tokens; up to 3 per task, shortest first, near-duplicates dropped. The teacher's thinking is removed: only the visible response is stored and scored.
  3. Pass 2. Tasks with no usable pass-1 response were retried with the gold answer in the prompt (template below) and their answers had to pass a leak filter that drops responses mentioning the provided answer.
  4. Scoring. Each kept response was placed after the student's non-thinking chat template and scored by the bf16 teacher with Hugging Face transformers; pass-2 responses also under the privileged prompt.

System prompt

You solve every task by treating it as an optimization problem.

1. Objective: state precisely what a correct or best answer must achieve, and how you would tell a better answer from a worse one.
2. Variables and constraints: identify what you are free to choose and what every valid answer must satisfy.
3. Search: pick the most efficient strategy for this structure, explore candidates, and discard any that break a constraint or score worse than one you already have.
4. Verify: check your best candidate against the objective and every constraint, and repair it if a check fails.
5. Generalize: in one sentence, name the principle that made the solution work and would carry over to similar tasks.

Write this reasoning out in your answer, one step after another, and keep it concise. Then give the final answer in the format the task asks for.

Pass-2 teacher template

Here is a task together with a correct answer to it.

Task:
{prompt}

Correct answer:
{privileged_context}

Solve the task again from scratch as an efficient optimization problem. Define the objective, the variables and the constraints, choose a search strategy that suits the structure of the task, and reason your way to the answer step by step, in a way that would carry over to similar tasks. Verify the result before stating it.

The correct answer is shown only so that your reasoning ends in the right place. Write as if you had never seen it: do not mention, quote or hint at it, do not work backwards from it, and do not state any result before you have derived it. Write out the entire reasoning in your answer, then the final answer in the format the task asks for.

Notes and caveats

  • Travelling salesman, large instances, are missing. Generation was stopped before the last 6 of 86 shards (2,228 tasks: 2,084 TSP with 9–10 cities and 144 with 7–8). On 7–8 city instances the teacher solved only 4–7% of first-pass attempts, so those tasks are mostly pass-2 rescues; filter with pass == 1 (or exclude_pass2 in the training code) if you want only answers the teacher found on its own.
  • float16 log-probabilities. top_logprobs are stored in half precision, so the tail mass computed from them can round to a tiny negative number.
  • Evaluation data is not included. The frozen evaluation and validation splits (held-out MATH-500, NL4OPT, MMLU-Pro, held-out puzzle types and scheduling) were never used for generation.
  • Licenses. The Qwen models are Apache 2.0, reasoning-gym is Apache 2.0, GSM8K is MIT and MATH is MIT. Check the licenses of the source datasets before redistributing.

Usage

from datasets import load_dataset

scores = load_dataset("IParraMartin/optimizer-distillation", "scores", split="train")
responses = load_dataset("IParraMartin/optimizer-distillation", "responses", split="train")
Downloads last month
68