Qwen3-4B-Opus-Reasoning-SFT
Trained with
Clawpathy using the Tinker platform.
Training Details
| Parameter | Value |
|---|
| Base model | Qwen/Qwen3-4B-Instruct-2507 |
| Method | Supervised Fine-Tuning |
| Dataset | nohurry/Opus-4.6-Reasoning-3000x-filtered |
| LoRA rank | 32 |
| Learning rate | 0.00018 |
| Steps | 50 |
| Final loss | 1.236125 |
| Tinker sampler path | tinker://332646c7-9a78-5c49-a58c-3ca971b142ce:train:0/sampler_weights/clawpathy-c942fbe6-f6b-final |
| Tinker state path | tinker://332646c7-9a78-5c49-a58c-3ca971b142ce:train:0/weights/clawpathy-c942fbe6-f6b-state |
Evaluation
{
"job_id": "c942fbe6-f6b",
"model": "Trained: Qwen/Qwen3-4B-Instruct-2507 (job c942fbe6-f6b)",
"eval_type": "benchmark",
"benchmarks": [
"Dataset SFT-match (50)",
"MMLU (0-shot, 200)"
],
"benchmark_names": [
"Dataset SFT-match (50)",
"MMLU (0-shot, 200)"
],
"domain": "general",
"accuracy": 0.21,
"metrics": {
"dataset_sft_match/token_f1": 0.3774,
"dataset_sft_match/exact_match": 0.0,
"dataset_sft_match/evaluated": 50,
"dataset_sft_match/split": "train",
"dataset_sft_match/format": "instruction",
"dataset_sft_match/examples": [
{
"expected": "# Solution: Calculating Bus Rental and Toll Costs\n\n## Step 1: Find the total number of people going on the field trip\n\n- Fifth-grade students: 252\n- Teachers: 8\n- Total people:",
"output": "## Solution\n\nStep 1: Calculate the total number of people going on the trip\n\n$$252 \text{ students} + 8 \text{ teachers} = 260 \text{ people}$$\n\nStep 2: Determine how many bu",
"token_f1": 0.3976,
"exact": false
},
{
"expected": "## Missing Problem Description\n\nYour request appears to be incomplete. You've provided the flushing instructions for an interactive problem, but the actual problem descriptio",
"output": "I notice you've provided instructions for an interactive problem, but you haven't actually included the problem statement or the specific question you'd like me to solve.\n\nCould yo",
"token_f1": 0.2332,
"exact": false
},
{
"expected": "Answer: no\n\nExplanation:\n\nThe premise states that "Thorn thought the same thing." For this statement to make sense, there must be another person (or people) who had this th",
"output": "## Solution\n\nStep 1: Analyze the premise\n- "Apparently Thorn thought the same thing" indicates that Thorn's thoughts aligned with someone else's thoughts\n- The word "same" impl",
"token_f1": 0.4814,
"exact": false
},
{
"expected": "## Solution\n\nThe question asks for automated software specifically designed to create energy profile diagrams.\n\n### Analysis of Options:\n\n| Option | Software | Type | Automated",
"output": "# Automated Software for Energy Profile Diagrams\n\n## Analysis of Each Option\n\nA: TikZ - This is a LaTeX package for creating graphics. While it can produce energy profile diagr",
"token_f1": 0.391,
"exact": false
},
{
"expected": "No, these sentences are not paraphrases of each other. \n\nA paraphrase would convey the same meaning or information using different wording. However, these two sentences describe en",
"output": "## Analysis\n\nLet me compare these two sentences to determine if they are paraphrases.\n\nSentence 1: "In Ventura County, the number of closings and advisories dropped by 73 perce",
"token_f1": 0.3219,
"exact": false
},
{
"expected": "The answer is C (contradiction).\n\nThe two sentences directly contradict each other. Sentence 1 explicitly states that the woman is wearing a WHITE dress while riding a bicycle. How",
"output": "# Analysis of Sentence Agreement\n\n## Step 1: Identify the key information in each sentence\n\nSentence 1: "A woman wearing a white dress riding a bicycle."\n- Subject: A woman\n- C",
"token_f1": 0.3739,
"exact": false
},
{
"expected": "# Article Based on Provided Highlights\n\n## Clinton to Suspend Presidential Campaign Saturday, Sources Say\n\nIn a significant development in the Democratic presidential race, sources",
"output": "# Clinton Suspends Presidential Campaign Amid Democratic Unity Efforts\n\n## Key Developments in the 2016 Democratic Primary Race\n\nWASHINGTON, D.C. \u2014 Democratic presidential nomi",
"token_f1": 0.4765,
"exact": false
},
{
"expected": "# Problem Analysis\n\nBased on the example provided, I can identify this as a modular arithmetic transformation problem.\n\n## Problem Statement (Reconstructed)\n\nGiven:\n- A modulus",
"output": "I notice you've provided an example but haven't actually asked a question. Let me analyze what this example demonstrates:\n\n## What this example shows:\n\nProblem Type: This appea",
"token_f1": 0.1667,
"exact": false
},
{
"expected": "It appears you've provided the output format for a competitive programming problem, but I'm missing the actual problem statement. To give you a detailed solution with verification,",
"output": "I notice that your message appears to be a partial problem statement, but it's incomplete. You've provided:\n\n- A conditional structure: "Otherwise, print an integer m..."\n- A descr",
"token_f1": 0.4075,
"exact": false
},
{
"expected": "Based on the context provided, the answer is:\n\nAll federal funds, including grants, loans, unions and employers who accepted taxpayer funds, and other forms of financial assist",
"output": "# Answer Extraction\n\n## Step-by-Step Analysis\n\nStep 1: Identify the question\n- "What was forced to comply?"\n\nStep 2: Locate relevant information in the context\n- The key se",
"token_f1": 0.5455,
"exact": false
}
],
"mmlu/accuracy": 0.21,
"mmlu/evaluated": 200,
"mmlu/correct": 42,
"mmlu/split": "validation",
"mmlu/limit": 200,
"mmlu/examples": [
{
"question": "The cyclic subgroup of Z_24 generated by 18 has order",
"gold": "A",
"pred": "A",
"raw": "# Solution\n\n## Step 1: Understanding the Problem\n\nWe need to find the order of t",
"correct": true
},
{
"question": "Find the order of the factor group Z_6/<3>.",
"gold": "B",
"pred": "C",
"raw": "# Solution: Order of the Factor Group Z_6/<3>\n\n## Step 1: Understand the problem",
"correct": false
},
{
"question": "Statement 1 | A permutation that is a product of m even permutations and n odd permutations is an even permutation if and only if n is even. Statement 2 | Every",
"gold": "A",
"pred": "A",
"raw": "# Solution\n\n## Statement 1 Analysis\n\nStatement 1: A permutation that is a pr",
"correct": true
},
{
"question": "Find the order of the factor group (Z_4 x Z_12)/(<2> x <2>)",
"gold": "C",
"pred": "A",
"raw": "# Solution\n\n## Step 1: Understand the problem\nWe need to find the order of the f",
"correct": false
},
{
"question": "Find the maximum possible order for some element of Z_4 x Z_6.",
"gold": "C",
"pred": "A",
"raw": "# Finding the Maximum Order in Z_4 \u00d7 Z_6\n\n## Step 1: Understanding the Problem\n\n",
"correct": false
},
{
"question": "Statement 1 | The symmetric group S_3 is cyclic. Statement 2 | Every group is isomorphic to some group of permutations.",
"gold": "D",
"pred": "A",
"raw": "# Solution\n\n## Statement 1: "The symmetric group S_3 is cyclic."\n\nAnalysis:\n",
"correct": false
},
{
"question": "Statement 1 | If a and b are elements of finite order in an Abelian group, then |ab| is the lcm (|a|,|b|). Statement 2 | If g is a group element and g^n = e, th",
"gold": "B",
"pred": "A",
"raw": "# Verification of Group Theory Statements\n\n## Statement 1 Analysis\nStatement 1",
"correct": false
},
{
"question": "Statement 1 | If f is a homomorphism from G to K and H is normal in G then f(H) is normal in K. Statement 2 | If f is a homomorphism from G to a group and H is ",
"gold": "D",
"pred": "A",
"raw": "# Solution\n\n## Statement 1 Analysis\nIf f is a homomorphism from G to K and H i",
"correct": false
},
{
"question": "Find the maximum possible order for an element of S_n for n = 7.",
"gold": "B",
"pred": "A",
"raw": "# Solution: Finding the Maximum Order of an Element in S\u2087\n\n## Step 1: Understand",
"correct": false
},
{
"question": "Statement 1 | Every integral domain has a field of quotients. Statement 2 | A polynomial of degree n over a ring can have at most n zeros counting multiplicity.",
"gold": "C",
"pred": "A",
"raw": "# Solution\n\n## Statement 1: Every integral domain has a field of quotients\n\n**TR",
"correct": false
}
]
},
"results": [
{
"benchmark": "dataset_sft_match/token_f1",
"score": 0.3774
},
{
"benchmark": "dataset_sft_match/exact_match",
"score": 0.0
},
{
"benchmark": "dataset_sft_match/evaluated",
"score": 50.0
},
{
"benchmark": "mmlu/accuracy",
"score": 0.21
},
{
"benchmark": "mmlu/evaluated",
"score": 200.0
},
{
"benchmark": "mmlu/correct",
"score": 42.0
},
{
"benchmark": "mmlu/limit",
"score": 200.0
}
],
"code_eval": null,
"stored_eval": null,
"dataset_id": "nohurry/Opus-4.6-Reasoning-3000x-filtered",
"status": "completed",
"requested_eval_tasks": [],
"executed_eval_tasks": [
"dataset_sft_match",
"internal/mmlu_0_shot"
],
"task_errors": {},
"is_proxy": false,
"is_complete": true
}