question stringclasses 5 values | metadata stringclasses 5 values | task_source stringclasses 1 value | formatted_prompt listlengths 2 2 | responses_by_sample listlengths 2 2 | tokens_by_sample listlengths 2 2 | num_rounds_by_sample listlengths 2 2 | continuation_prompts_by_sample listlengths 2 2 | continuation_text_by_sample listlengths 2 2 | strategy_type stringclasses 1 value | termination_reason_by_sample listlengths 2 2 | correct_round_by_sample listlengths 2 2 | smart_termination_enabled bool 1 class | retry_count_by_sample listlengths 2 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Using ALL of the numbers 76, 55, 5, 4, create an expression that equals 122. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [76, 55, 5, 4], "target": 122, "operators_used": ["+", "-", "-"], "expression": "76 + 55 - 5 - 4"} | countdown | [
{
"content": "You are a helpful assistant that solves math problems by reasoning step-by-step. Show your work clearly and provide complete solutions. You MUST use ALL of the given numbers exactly once in your solution. When you have a final answer, wrap it in <answer></answer> tags.",
"role": "system"
},
... | [
[
"I need to get 122.\nAvailable numbers: [76, 55, 5, 4]\n\nWhat if I compute 76 + 5?\n76 + 5 = 81.\n\nLet me try 81 + 4.\nThat gives 81 + 4 = 85.\nGot 85, need 37 more.\n\nTrying 55 * 85.\n55 * 85 = 4675.\n\nWait, let me try a different approach.\n\nThis path doesn't lead to 122. Backtracking...\n\nThis path d... | [
174,
57
] | [
1,
1
] | [
[],
[]
] | [
[],
[]
] | dfs_baseline | [
"budget_reached",
"answer_correct"
] | [
null,
0
] | true | [
0,
0
] |
Using ALL of the numbers 45, 78, 34, 4, create an expression that equals 886. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [45, 78, 34, 4], "target": 886, "operators_used": ["*", "+", "/"], "expression": "45 * 78 + 34 / 4"} | countdown | [
{
"content": "You are a helpful assistant that solves math problems by reasoning step-by-step. Show your work clearly and provide complete solutions. You MUST use ALL of the given numbers exactly once in your solution. When you have a final answer, wrap it in <answer></answer> tags.",
"role": "system"
},
... | [
[
"Target: 886. Let me think...\nAvailable numbers: [45, 78, 34, 4]\n\n78 * 4 = 312.\nNow I have 312. I need 574 more to reach 886.\n\nTrying 45 + 312.\nWe get 45 + 312 = 357.\n\nTrying 34 + 357.\nSo 34 + 357 = 391.\n\nThat doesn't work. Let me try something else.\n\nThat doesn't work. Let me try something else... | [
119,
373
] | [
1,
1
] | [
[],
[]
] | [
[],
[]
] | dfs_baseline | [
"budget_reached",
"budget_reached"
] | [
null,
null
] | true | [
0,
0
] |
Using ALL of the numbers 91, 9, 6, 85, create an expression that equals 577. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [91, 9, 6, 85], "target": 577, "operators_used": ["-", "*", "+"], "expression": "91 - 9 * 6 + 85"} | countdown | [
{
"content": "You are a helpful assistant that solves math problems by reasoning step-by-step. Show your work clearly and provide complete solutions. You MUST use ALL of the given numbers exactly once in your solution. When you have a final answer, wrap it in <answer></answer> tags.",
"role": "system"
},
... | [
[
"I need to get 577.\nAvailable numbers: [91, 9, 6, 85]\n\nThat gives 91 * 6 = 546.\n\n91 + 510 = 601.\nThat's 601. Still need 24.\n\nTrying 601 - 9.\nSo 601 - 9 = 592.\n\nThat doesn't work. Let me try something else.\n\nSo 9 + 546 = 555.\nThat's 555. Still need 22.\n\n85 + 555 = 640.\n\nWait, let me try a dif... | [
171,
93
] | [
1,
1
] | [
[],
[]
] | [
[],
[]
] | dfs_baseline | [
"budget_reached",
"budget_reached"
] | [
null,
null
] | true | [
0,
0
] |
Using ALL of the numbers 66, 64, 12, 1, create an expression that equals 142. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [66, 64, 12, 1], "target": 142, "operators_used": ["+", "+", "/"], "expression": "66 + 64 + 12 / 1"} | countdown | [
{
"content": "You are a helpful assistant that solves math problems by reasoning step-by-step. Show your work clearly and provide complete solutions. You MUST use ALL of the given numbers exactly once in your solution. When you have a final answer, wrap it in <answer></answer> tags.",
"role": "system"
},
... | [
[
"Let's find a way to make 142.\nAvailable numbers: [66, 64, 12, 1]\n\nThat gives 66 + 12 = 78.\n\nWhat if I compute 64 + 78?\nThat gives 64 + 78 = 142.\nGot it: 142 = 142.\n\n<answer>142</answer>"
],
[
"I need to get 142.\nAvailable numbers: [66, 64, 12, 1]\n\nSo 66 + 64 = 130.\n\nSo 12 + 130 = 142.\n... | [
52,
37
] | [
1,
1
] | [
[],
[]
] | [
[],
[]
] | dfs_baseline | [
"answer_correct",
"answer_correct"
] | [
0,
0
] | true | [
0,
0
] |
Using ALL of the numbers 15, 47, 40, 31, create an expression that equals 696. You must use each number exactly once. You may use +, -, *, /. | {"numbers": [15, 47, 40, 31], "target": 696, "operators_used": ["*", "-", "+"], "expression": "15 * 47 - 40 + 31"} | countdown | [
{
"content": "You are a helpful assistant that solves math problems by reasoning step-by-step. Show your work clearly and provide complete solutions. You MUST use ALL of the given numbers exactly once in your solution. When you have a final answer, wrap it in <answer></answer> tags.",
"role": "system"
},
... | [
[
"Target: 696. Let me think...\nAvailable numbers: [15, 47, 40, 31]\n\nWhat if I compute 15 * 47?\nWe get 15 * 47 = 705.\n\nLet me try 705 - 31.\nThat gives 705 - 31 = 674.\nNow I have 674. I need 22 more to reach 696.\n\nTrying 40 + 674.\nWe get 40 + 674 = 714.\n\nThat doesn't work. Let me try something else.... | [
115,
75
] | [
1,
1
] | [
[],
[]
] | [
[],
[]
] | dfs_baseline | [
"budget_reached",
"budget_reached"
] | [
null,
null
] | true | [
0,
0
] |
test-dfs-baseline
Countdown reasoning with dfs_baseline strategy to 2000 tokens
Dataset Info
- Rows: 5
- Columns: 14
Columns
| Column | Type | Description |
|---|---|---|
| question | Value('string') | The countdown problem statement |
| metadata | Value('string') | JSON with target number, available numbers, and solution |
| task_source | Value('string') | Task type identifier (countdown) |
| formatted_prompt | List({'content': Value('string'), 'role': Value('string')}) | Initial prompt messages sent to the model |
| responses_by_sample | List(List(Value('string'))) | Full model response [sample_idx][round_idx] = response string |
| tokens_by_sample | List(Value('int64')) | Total token count for each sample (based on continuation_text) |
| num_rounds_by_sample | List(Value('int64')) | Number of continuation rounds for each sample |
| continuation_prompts_by_sample | List(List(Value('null'))) | Continuation prompts used [sample_idx][round_idx] = prompt string |
| continuation_text_by_sample | List(List(Value('null'))) | Truncated text used for continuation [sample_idx][round_idx]. For budget_forcing: truncated at first tag. For other strategies: same as responses_by_sample. |
| strategy_type | Value('string') | Prompting strategy: 'iterative', 'budget_forcing', or 'random_noise' |
| termination_reason_by_sample | List(Value('string')) | Why each sample terminated: 'max_rounds_reached', 'budget_reached', or 'answer_correct' |
| correct_round_by_sample | List(Value('int64')) | Round number where correct answer was first found (None if never correct) |
| smart_termination_enabled | Value('bool') | Whether smart termination was enabled for this run |
| retry_count_by_sample | List(Value('int64')) | Total retry attempts per sample due to empty or too-short responses (0 if no retries needed) |
Generation Parameters
{
"script_name": "stage_01_generate_reasoning.py",
"model": "dfs_baseline",
"hyperparameters": {
"temperature": 0.7,
"max_tokens_per_round": 2000,
"target_total_tokens": 2000,
"samples": 2,
"strategy": "dfs_baseline",
"max_rounds": null,
"min_tokens": null,
"max_retries_per_round": -1
},
"input_datasets": [],
"description": "Countdown reasoning with dfs_baseline strategy to 2000 tokens",
"experiment": "degen_test_1_countdown",
"strategy": "dfs_baseline",
"arg_count": 4,
"seed": 42,
"smart_termination": false,
"max_rounds": null,
"min_tokens": null,
"inference_backend": "curator",
"status": "completed",
"completed_at": "2026-01-01T23:15:22.836847"
}
Experiment Documentation
For complete experiment details, see https://github.com/TAUR-Lab/Reasoning-Horizons/tree/main/experiments/degen_test_1_countdown
Usage
from datasets import load_dataset
dataset = load_dataset("reasoning-degeneration-dev/test-dfs-baseline", split="train")
print(f"Loaded {len(dataset)} rows")
This dataset is tracked in reasoning-degeneration-dev/PROJECT-MANIFEST
- Downloads last month
- 5