question stringclasses 3
values | metadata stringclasses 3
values | task_source stringclasses 1
value | formatted_prompt listlengths 2 2 | response listlengths 1 1 | response__extractions listlengths 1 1 | response__extractions_metadata listlengths 1 1 | eval_correct listlengths 1 1 | strategy_compliance int64 1 1 | compliance_reasoning stringclasses 3
values | compliance_elements_followed listlengths 0 0 | compliance_elements_missed listlengths 8 8 |
|---|---|---|---|---|---|---|---|---|---|---|---|
Using ALL of the numbers 18, 3, 14, 87, 95, 70, 12, 76, create an expression that equals 196. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [18, 3, 14, 87, 95, 70, 12, 76], "target": 196, "operators_used": ["/", "+", "+", "+", "-", "-", "+"], "expression": "18 / 3 + 14 + 87 + 95 - 70 - 12 + 76"} | countdown | [
{
"content": "You are a math problem solver. Solve the given arithmetic problem step by step. Show your work and then provide your final answer inside <answer> tags.\n\nFor example: <answer>10 + 5 * 2 = 20</answer>\n\nThe answer should be a single arithmetic expression using the given numbers that equals the ta... | [
"Okay, let's see. I need to use all these numbers: 18, 3, 14, 87, 95, 70, 12, 76 to make 196. Hmm, that's eight numbers. Each exactly once, using +, -, *, /. Alright, so first thought: maybe some combination of additions and subtractions, but probably with some multiplication or division to get the right total. Let... | [
[]
] | [
[]
] | [
[]
] | 1 | The model completely ignores the prescribed strategy and relies on unstructured exploration and ad-hoc reasoning throughout its response. It does not follow any of the systematic steps outlined in the strategy, such as listing available numbers, branching systematically, or marking dead ends. | [] | [
"Lists all available numbers and target at the start (INVENTORY step)",
"Picks pairs of numbers and tries operations systematically (BRANCH step)",
"Tracks the available number set after each operation (shows remaining numbers)",
"Uses explicit tree structure (branches, levels, indentation, or labeling)",
"... |
Using ALL of the numbers 77, 42, 63, 3, 15, 45, 40, 31, create an expression that equals 176. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [77, 42, 63, 3, 15, 45, 40, 31], "target": 176, "operators_used": ["+", "*", "+", "+", "/", "+", "-"], "expression": "77 + 42 * 63 + 3 + 15 / 45 + 40 - 31"} | countdown | [
{
"content": "You are a math problem solver. Solve the given arithmetic problem step by step. Show your work and then provide your final answer inside <answer> tags.\n\nFor example: <answer>10 + 5 * 2 = 20</answer>\n\nThe answer should be a single arithmetic expression using the given numbers that equals the ta... | [
"Okay, let's see. I need to use all the numbers 77, 42, 63, 3, 15, 45, 40, 31 to make 176. Hmm, that's a bit challenging because there are eight numbers, and I have to combine them with +, -, *, / to get exactly 176. Let me start by thinking about possible ways to break this down.\n\nFirst, maybe I should consider ... | [
[]
] | [
[]
] | [
[]
] | 1 | The model completely ignores the prescribed strategy and relies on unstructured exploration and ad-hoc reasoning. It does not follow any of the systematic steps outlined in the strategy, such as listing available numbers, branching systematically, or marking dead ends. | [] | [
"Lists all available numbers and target at the start (INVENTORY step)",
"Picks pairs of numbers and tries operations systematically (BRANCH step)",
"Tracks the available number set after each operation (shows remaining numbers)",
"Uses explicit tree structure (branches, levels, indentation, or labeling)",
"... |
Using ALL of the numbers 53, 60, 1, 7, 87, 84, 83, 13, create an expression that equals 698. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [53, 60, 1, 7, 87, 84, 83, 13], "target": 698, "operators_used": ["+", "/", "*", "+", "-", "-", "-"], "expression": "53 + 60 / 1 * 7 + 87 - 84 - 83 - 13"} | countdown | [
{
"content": "You are a math problem solver. Solve the given arithmetic problem step by step. Show your work and then provide your final answer inside <answer> tags.\n\nFor example: <answer>10 + 5 * 2 = 20</answer>\n\nThe answer should be a single arithmetic expression using the given numbers that equals the ta... | [
"Okay, let's see. I need to use all these numbers: 53, 60, 1, 7, 87, 84, 83, 13 to make 698. Hmm, that's a bit tricky. Let me start by thinking about how to combine these numbers. Maybe start by adding some of the larger numbers and see how close I can get.\n\nFirst, let's list all the numbers again to make sure I ... | [
[]
] | [
[]
] | [
[]
] | 1 | The model completely ignores the prescribed strategy and relies on a default reasoning approach, primarily involving ad-hoc guessing and estimation. It does not systematically list available numbers or follow the structured tree search format, leading to a lack of compliance with the strategy. | [] | [
"Lists all available numbers and target at the start (INVENTORY step)",
"Picks pairs of numbers and tries operations systematically (BRANCH step)",
"Tracks the available number set after each operation (shows remaining numbers)",
"Uses explicit tree structure (branches, levels, indentation, or labeling)",
"... |
t1-strategy-countdown-baseline-together_ai-qwen-qwen3-next-80b-a3b-thin-b00e6e41
Strategy compliance evaluation on countdown arithmetic — baseline variant.
Model received no strategy instruction (baseline control). Judge still scores against tree search rubric. Compliance is scored by an LLM judge (1-5 Likert).
Results
| Metric | Value |
|---|---|
| pass@1 | 0.0000 |
| Strategy compliance (mean) | 1.00 |
| Strategy compliance (min) | 1 |
| Strategy compliance (max) | 1 |
| Total problems | 3 |
Compliance Distribution
| Score | Count | Meaning |
|---|---|---|
| 1 | 3 | Non-compliant |
| 2 | 0 | Minimal |
| 3 | 0 | Partial |
| 4 | 0 | Mostly compliant |
| 5 | 0 | Fully compliant |
Details
| Parameter | Value |
|---|---|
| Model | together_ai/Qwen/Qwen3-Next-80B-A3B-Thinking |
| Thinking | True |
| Temperature | 0.7 |
| Max tokens | 16384 |
| Judge model | openai/gpt-4o-mini |
| Num examples | 3 |
| Num args | 8 |
| Variant | baseline |
Usage
from datasets import load_dataset
ds = load_dataset("reasoning-degeneration-dev/t1-strategy-countdown-baseline-together_ai-qwen-qwen3-next-80b-a3b-thin-b00e6e41", split="train")
scores = ds["strategy_compliance"]
print(f"Mean compliance: {sum(scores) / len(scores):.2f}")
Tracked in reasoning-degeneration-dev/PROJECT-MANIFEST
- Downloads last month
- 14