narrative string | question string | choices string | answer_index int64 | answer_choice string | task_source string | formatted_prompt list | response list | formatted_prompt__metadata list | response__metadata list | eval_result list | eval_result__metadata list | _judge_prompt__metadata list | _judge_response__metadata list | strategy_compliance int64 | compliance_reasoning string | compliance_elements_followed list | compliance_elements_missed list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
In the tranquil solitude of a secluded chalet, Clyde meets an untimely death from snake venom; a case which sees the sharp mind of Detective Winston matching wits with two elusive suspects, Mack and Taylor.
Detective Winston was looking at the tailor-stitched files bundled together with papers on a desk in his office.... | Who is the most likely murderer? | ['Mack', 'Taylor'] | 0 | Mack | musr | [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"content": "In the tranquil solitude of a secluded chalet, Clyde meets an untimely death from snake venom; a case which sees the sharp mind of Detective Winston matching wits with two ... | [
"Let’s examine each suspect against the three classic requirements: \n- Means – did they possess or could they obtain snake venom? \n- Motive – did they have a strong reason to want Clyde dead? \n- Opportunity – were they able to be at the secluded chalet at the relevant time?\n\n--------------------------------... | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"raw_prompt": [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"conten... | [
{
"cost": 0.0026750000000000003,
"finish_reason": "stop",
"input_tokens": 1340,
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"output_tokens": 445,
"raw_response": "{\"id\": \"oYMavs3-zqrih-9d1c3f747d97134a\", \"created\": 1771739875, \"model\": \"together_ai/moonshotai/Kimi-K2-In... | [
true
] | [
{
"correct_answer_index": 0,
"correct_answer_number": 1,
"correct_answer_text": "Mack",
"extracted_number": 1,
"extracted_text": "1",
"extractor": "musr",
"matched_by": "number",
"num_choices": 2,
"raw_line": "ANSWER: 1"
}
] | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "gpt-4o-mini-2024-07-18",
"raw_prompt": [
{
"content": "You are an expert evaluator assessing whether a language model followed a prescribed reasoning strategy. You will be given:\n\n1. The STRATEGY INSTRUCTION that... | [
{
"cost": 0.00030884999999999994,
"finish_reason": "stop",
"input_tokens": 1215,
"model": "gpt-4o-mini-2024-07-18",
"output_tokens": 211,
"raw_response": "{\"id\": \"chatcmpl-DBwlZgqXihqN5pnm3h1OXtMz0BX9t\", \"created\": 1771739877, \"model\": \"gpt-4o-mini-2024-07-18\", \"object\": \"chat.c... | 5 | The model fully complies with the prescribed strategy by systematically listing all suspects, creating a detailed matrix, and filling it with specific evidence from the narrative. Each cell is marked appropriately, and the analysis leads to a clear conclusion based on the matrix, demonstrating a thorough application of... | [
"Lists all suspects at the start (SUSPECT LIST step)",
"Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)",
"Fills matrix cells with specific evidence from the narrative",
"Marks each cell with a status (CONFIRMED/DENIED/UNKNOWN or equivalent)",
"Processes ALL suspects sy... | [] |
In the dim light of an abandoned factory, Kaylee's lifeless body lay with the cold steel of a pickaxe lodged within, thrusting Detective Winston into a dark game of deception with suspects Penelope and Francis.
Detective Winston sat at the precinct and scrutinized the last seen diary of Kaylee, a murder victim. The di... | Who is the most likely murderer? | ['Penelope', 'Francis'] | 0 | Penelope | musr | [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"content": "In the dim light of an abandoned factory, Kaylee's lifeless body lay with the cold steel of a pickaxe lodged within, thrusting Detective Winston into a dark game of decepti... | [
"Let’s weigh each suspect against the three indispensable pillars of a murder case: motive, means, and opportunity.\n\nPENELOPE \n1. Motive – Crystal-clear. Kaylee’s diary shows she was about to expose Penelope’s hidden criminal record to the parents who employ her as a nanny. Loss of livelihood and social ruin is... | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"raw_prompt": [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"conten... | [
{
"cost": 0.00253,
"finish_reason": "stop",
"input_tokens": 1393,
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"output_tokens": 379,
"raw_response": "{\"id\": \"oYMavnD-4qdXgU-9d1c3f746fa90f55\", \"created\": 1771739875, \"model\": \"together_ai/moonshotai/Kimi-K2-Instruct-0905\"... | [
true
] | [
{
"correct_answer_index": 0,
"correct_answer_number": 1,
"correct_answer_text": "Penelope",
"extracted_number": 1,
"extracted_text": "1",
"extractor": "musr",
"matched_by": "number",
"num_choices": 2,
"raw_line": "ANSWER: 1"
}
] | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "gpt-4o-mini-2024-07-18",
"raw_prompt": [
{
"content": "You are an expert evaluator assessing whether a language model followed a prescribed reasoning strategy. You will be given:\n\n1. The STRATEGY INSTRUCTION that... | [
{
"cost": 0.00028335000000000003,
"finish_reason": "stop",
"input_tokens": 1161,
"model": "gpt-4o-mini-2024-07-18",
"output_tokens": 182,
"raw_response": "{\"id\": \"chatcmpl-DBwlZegz1GeTkoAM2PvCfbLPvm7DZ\", \"created\": 1771739877, \"model\": \"gpt-4o-mini-2024-07-18\", \"object\": \"chat.c... | 3 | The model partially follows the strategy by discussing suspects and their motives, means, and opportunities, but it does not create an explicit matrix as required. While it provides evidence for each suspect, it lacks the structured table format and does not mark the cells with statuses like CONFIRMED or DENIED, which ... | [
"Lists all suspects at the start",
"Cites specific narrative evidence",
"Processes ALL suspects systematically"
] | [
"Creates an explicit matrix/table with suspects vs criteria",
"Fills matrix cells with specific evidence from the narrative",
"Marks each cell with a status (CONFIRMED/DENIED/UNKNOWN or equivalent)",
"Uses the matrix to derive the answer (elimination logic visible)",
"Does not skip to the answer before buil... |
In a shocking twist of events, Charlotte is found lifeless in her kitchen, felled by a stun gun, and it's up to seasoned Detective Winston to unravel the mystery, with prime suspects being Letti and Deborah.
Winston lit up a fresh cigarette, the inky night a silent observer as he reviewed Charlotte's file. It wasn't ... | Who is the most likely murderer? | ['Letti', 'Deborah'] | 0 | Letti | musr | [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"content": "In a shocking twist of events, Charlotte is found lifeless in her kitchen, felled by a stun gun, and it's up to seasoned Detective Winston to unravel the mystery, with prim... | [
"Step 1 – Establish the three required elements for each suspect \n\nA. Deborah \n- Means (weapon): nothing links her to a stun-gun purchase; the only stun-gun on the record was bought by Letti earlier that day. \n- Motive: overwhelming. Charlotte had beaten, terrorised and sent scores of death-threats to her so... | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"raw_prompt": [
{
"content": "You are a helpful assistant that will answer the questions given by the user.",
"role": "system"
},
{
"conten... | [
{
"cost": 0.002607,
"finish_reason": "stop",
"input_tokens": 1422,
"model": "together_ai/moonshotai/Kimi-K2-Instruct-0905",
"output_tokens": 395,
"raw_response": "{\"id\": \"oYMavq6-zqrih-9d1c3f746a3042c0\", \"created\": 1771739874, \"model\": \"together_ai/moonshotai/Kimi-K2-Instruct-0905\"... | [
true
] | [
{
"correct_answer_index": 0,
"correct_answer_number": 1,
"correct_answer_text": "Letti",
"extracted_number": 1,
"extracted_text": "1",
"extractor": "musr",
"matched_by": "number",
"num_choices": 2,
"raw_line": "ANSWER: 1"
}
] | [
{
"endpoint_type": "chat",
"litellm_function": "acompletion",
"model": "gpt-4o-mini-2024-07-18",
"raw_prompt": [
{
"content": "You are an expert evaluator assessing whether a language model followed a prescribed reasoning strategy. You will be given:\n\n1. The STRATEGY INSTRUCTION that... | [
{
"cost": 0.0003006,
"finish_reason": "stop",
"input_tokens": 1172,
"model": "gpt-4o-mini-2024-07-18",
"output_tokens": 208,
"raw_response": "{\"id\": \"chatcmpl-DBwlZKUWyWdHRfd2Q98YhG3lhkAUC\", \"created\": 1771739877, \"model\": \"gpt-4o-mini-2024-07-18\", \"object\": \"chat.completion\", ... | 3 | The model partially follows the strategy by listing suspects and providing evidence for each, but it does not create an explicit matrix as required. While it analyzes the suspects, it skips the matrix-building step and does not mark the cells with statuses, which are crucial for compliance. | [
"Lists all suspects at the start (SUSPECT LIST step)",
"Cites specific narrative evidence (not just assertions)"
] | [
"Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)",
"Fills matrix cells with specific evidence from the narrative",
"Marks each cell with a status (CONFIRMED/DENIED/UNKNOWN or equivalent)",
"Processes ALL suspects systematically, not just likely ones",
"Uses the matrix t... |
"In the high-intensity world of a local fitness center, Tommy's exercise regime is abruptly terminat(...TRUNCATED) | Who is the most likely murderer? | ['Lucy', 'Jacqueline'] | 0 | Lucy | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Step-by-step check against the three legal elements:\n\n1. Lucy \n – Motive: Strong. Tommy wa(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.002264,"finish_reason":"stop","input_tokens":1358,"model":"together_ai/moonshotai/Kimi-K2(...TRUNCATED) | [
true
] | [{"correct_answer_index":0,"correct_answer_number":1,"correct_answer_text":"Lucy","extracted_number"(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.00027315,"finish_reason":"stop","input_tokens":1081,"model":"gpt-4o-mini-2024-07-18","out(...TRUNCATED) | 3 | "The model partially follows the strategy by listing suspects and providing some analysis of their m(...TRUNCATED) | [
"Lists all suspects at the start",
"Cites specific narrative evidence"
] | ["Creates an explicit matrix/table with suspects vs criteria","Fills matrix cells with specific evid(...TRUNCATED) |
"In a chilling turn of events, Rose is found lifeless in a car, killed by a vial of acid, leading De(...TRUNCATED) | Who is the most likely murderer? | ['Dexter', 'Daisy'] | 1 | Daisy | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Step-by-step check of the three elements for each suspect:\n\n1. Dexter \n – Means: YES. Coll(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.002309,"finish_reason":"stop","input_tokens":1367,"model":"together_ai/moonshotai/Kimi-K2(...TRUNCATED) | [
true
] | [{"correct_answer_index":1,"correct_answer_number":2,"correct_answer_text":"Daisy","extracted_number(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.00028514999999999997,"finish_reason":"stop","input_tokens":1093,"model":"gpt-4o-mini-2024(...TRUNCATED) | 4 | "The model mostly follows the prescribed strategy by listing suspects, analyzing each one systematic(...TRUNCATED) | ["Lists all suspects at the start (SUSPECT LIST step)","Fills matrix cells with specific evidence fr(...TRUNCATED) | ["Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)","Does not s(...TRUNCATED) |
"In the dense, dank heart of the rainforest, Skylar's life was brutally silenced with a lead pipe, l(...TRUNCATED) | Who is the most likely murderer? | ['Albert', 'Mya'] | 1 | Mya | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Step-by-step check of the three required elements for each suspect.\n\n1. Albert \n – Motive:(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.002369,"finish_reason":"stop","input_tokens":1466,"model":"together_ai/moonshotai/Kimi-K2(...TRUNCATED) | [
true
] | [{"correct_answer_index":1,"correct_answer_number":2,"correct_answer_text":"Mya","extracted_number":(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.00028485,"finish_reason":"stop","input_tokens":1083,"model":"gpt-4o-mini-2024-07-18","out(...TRUNCATED) | 4 | "The model mostly follows the prescribed strategy by listing suspects, analyzing their motives, mean(...TRUNCATED) | ["Lists all suspects at the start (SUSPECT LIST step)","Fills matrix cells with specific evidence fr(...TRUNCATED) | ["Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)","Marks each(...TRUNCATED) |
"In the hallowed silence of a synagogue, a shocking murder unfolds as Derek is impaled by a harpoon,(...TRUNCATED) | Who is the most likely murderer? | ['Hailey', 'Angela'] | 1 | Angela | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Step-by-step check of the three requirements for each suspect.\n\n1. MEANS – access to the murde(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.0030629999999999998,"finish_reason":"stop","input_tokens":1719,"model":"together_ai/moons(...TRUNCATED) | [
true
] | [{"correct_answer_index":1,"correct_answer_number":2,"correct_answer_text":"Angela","extracted_numbe(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.0003126,"finish_reason":"stop","input_tokens":1224,"model":"gpt-4o-mini-2024-07-18","outp(...TRUNCATED) | 4 | "The model mostly follows the prescribed strategy by systematically analyzing the suspects and provi(...TRUNCATED) | ["Lists all suspects at the start (SUSPECT LIST step)","Fills matrix cells with specific evidence fr(...TRUNCATED) | ["Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)","Marks each(...TRUNCATED) |
"In the haze of neon lights and the serving of a silent hand of fate, Timothy lies dead in a casino,(...TRUNCATED) | Who is the most likely murderer? | ['Harry', 'Rosemary'] | 1 | Rosemary | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Let’s weigh the two suspects against the three classic requirements: motive, means, opportunity.(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.002435,"finish_reason":"stop","input_tokens":1526,"model":"together_ai/moonshotai/Kimi-K2(...TRUNCATED) | [
false
] | [{"correct_answer_index":1,"correct_answer_number":2,"correct_answer_text":"Rosemary","extracted_num(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.000267,"finish_reason":"stop","input_tokens":1084,"model":"gpt-4o-mini-2024-07-18","outpu(...TRUNCATED) | 2 | "The model identifies suspects and discusses their motives, means, and opportunities, but it does no(...TRUNCATED) | [
"Lists all suspects at the start",
"Cites specific narrative evidence"
] | ["Creates an explicit matrix/table with suspects vs criteria","Fills matrix cells with specific evid(...TRUNCATED) |
"In the aftermath of a devastating home explosion that claims the life of Angela, seasoned Detective(...TRUNCATED) | Who is the most likely murderer? | ['Patrick', 'Alvin'] | 0 | Patrick | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Let’s test each suspect against the three classic requirements: motive, means, and opportunity.\(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.00277,"finish_reason":"stop","input_tokens":1462,"model":"together_ai/moonshotai/Kimi-K2-(...TRUNCATED) | [
true
] | [{"correct_answer_index":0,"correct_answer_number":1,"correct_answer_text":"Patrick","extracted_numb(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.00028665,"finish_reason":"stop","input_tokens":1211,"model":"gpt-4o-mini-2024-07-18","out(...TRUNCATED) | 3 | "The model partially follows the strategy by discussing suspects and their motives, means, and oppor(...TRUNCATED) | ["Lists all suspects at the start","Cites specific narrative evidence","Processes ALL suspects syste(...TRUNCATED) | ["Creates an explicit matrix/table with suspects vs criteria","Fills matrix cells with specific evid(...TRUNCATED) |
"In the frost-kissed tranquility of a chalet, Malaysia meets a fatal end by a rifle's sinister call,(...TRUNCATED) | Who is the most likely murderer? | ['Letti', 'Evelyn'] | 1 | Evelyn | musr | [{"content":"You are a helpful assistant that will answer the questions given by the user.","role":"(...TRUNCATED) | ["Step 1 – Identify what must be proved \nFor either woman to be the shooter we need: \nA. Motiv(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"together_ai/moonshotai/Kimi-K2-In(...TRUNCATED) | [{"cost":0.0026509999999999997,"finish_reason":"stop","input_tokens":1283,"model":"together_ai/moons(...TRUNCATED) | [
true
] | [{"correct_answer_index":1,"correct_answer_number":2,"correct_answer_text":"Evelyn","extracted_numbe(...TRUNCATED) | [{"endpoint_type":"chat","litellm_function":"acompletion","model":"gpt-4o-mini-2024-07-18","raw_prom(...TRUNCATED) | [{"cost":0.00030435,"finish_reason":"stop","input_tokens":1233,"model":"gpt-4o-mini-2024-07-18","out(...TRUNCATED) | 4 | "The model mostly follows the prescribed strategy by identifying suspects, evaluating them systemati(...TRUNCATED) | ["Lists all suspects at the start (SUSPECT LIST step)","Fills matrix cells with specific evidence fr(...TRUNCATED) | ["Creates an explicit matrix/table with suspects vs criteria (MEANS/MOTIVE/OPPORTUNITY)","Does not s(...TRUNCATED) |
t1-strategy-musr-baseline-kimi-k2-instruct-musr-base
Strategy compliance evaluation on MuSR murder mysteries — baseline variant.
Model received standard MuSR cot+ prompt (baseline control). Judge still scores against evidence matrix rubric. Compliance is scored by an LLM judge (1-5 Likert) against the Evidence Matrix rubric.
Results
| Metric | Value |
|---|---|
| pass@1 | 0.9000 |
| Strategy compliance (mean) | 3.50 |
| Strategy compliance (min) | 2 |
| Strategy compliance (max) | 5 |
| Total problems | 10 |
Compliance Distribution
| Score | Count | Meaning |
|---|---|---|
| 1 | 0 | Non-compliant |
| 2 | 1 | Minimal |
| 3 | 4 | Partial |
| 4 | 4 | Mostly compliant |
| 5 | 1 | Fully compliant |
Details
| Parameter | Value |
|---|---|
| Model | together_ai/moonshotai/Kimi-K2-Instruct-0905 |
| Thinking | False |
| Temperature | 0.7 |
| Max tokens | 32768 |
| Judge model | openai/gpt-4o-mini |
| Num examples | 10 |
| Variant | baseline |
Usage
from datasets import load_dataset
ds = load_dataset("reasoning-degeneration-dev/t1-strategy-musr-baseline-kimi-k2-instruct-musr-base", 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
- 11