Datasets:
File size: 3,283 Bytes
425d315 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 | version = 1
[study]
id = "position_and_anonymity"
title = "Position (MT-Bench)"
description = "Measure response-position effects, always judged anonymously, on the full MT-Bench human-judgment dataset (turn 1, majority-vote across annotators)."
dataset = "data/mt_bench_prepared.parquet"
dataset_rows = 1199
dataset_sha256 = "f02359af006be11bda6903914ee10477716044623b37de2def944b9162d03c63"
api_config = "../configs/apis.json"
judge_config = "../configs/judges.json"
judges = [
# OpenAI
"openai-gpt-4.1-nano-batch-direct", # paid, cheapest OpenAI batch model
"ollama-gpt-oss-20b-low", # free, open-weight OpenAI model (reasoning cannot be disabled on this Ollama build)
# Anthropic (no open-weight Claude exists, so batch only)
"anthropic-claude-haiku-4.5-batch-direct",
# Google: Gemini (paid batch) + Gemma size ladder (free, local)
"google-gemini-3.1-flash-lite-batch-minimal",
"ollama-gemma-3n-e4b-direct", # smaller/edge
"ollama-gemma-3-27b-direct", # larger dense
# Meta
"ollama-llama-3.3-70b-direct",
# Alibaba (Qwen): reasoning toggle at fixed size
"ollama-qwen3-30b-a3b-direct",
"ollama-qwen3-30b-a3b-low",
# Mistral AI
"ollama-mistral-small-3.2-24b-direct",
# DeepSeek: full size ladder of R1 distillations (reasoning is always on
# for R1; there is no non-reasoning counterpart to pair against)
"ollama-deepseek-r1-1.5b-low",
"ollama-deepseek-r1-7b-low",
"ollama-deepseek-r1-8b-low",
"ollama-deepseek-r1-14b-low",
"ollama-deepseek-r1-32b-low",
"ollama-deepseek-r1-70b-low",
# Zhipu AI / Z.ai (GLM): older/smaller generation, plus a same-weights
# reasoning on/off pair on the current generation
"ollama-glm4-9b-direct",
"ollama-glm-4.7-flash-direct",
"ollama-glm-4.7-flash-thinking",
# Microsoft
"ollama-phi4-14b-direct",
# Stability AI
"ollama-stablelm2-12b-direct",
# UC Berkeley (NEST lab) — closest current proxy for koala-13b's origin lab
"ollama-starling-lm-7b-direct",
]
system_prompt = "prompts/system.txt"
user_prompt_template = "prompts/user.txt"
results_dir = "results"
[collection]
num_queries_per_record = 1
start_index = 0
max_concurrency = 1
timeout = 120
max_retries = 3
[[factors]]
id = "position"
label = "Response position"
description = "Whether the dataset responses retain their original A/B order or are swapped."
parameter = "response_order_default"
[[factors.levels]]
id = "original"
label = "Original order"
description = "Show response A first and response B second."
value = true
[[factors.levels]]
id = "swapped"
label = "Swapped order"
description = "Show response B first and response A second."
value = false
[[factors]]
id = "anonymity"
label = "Model-label visibility"
description = "Fixed: assistant labels always hide the source model names."
parameter = "is_anonymous"
[[factors.levels]]
id = "anonymous"
label = "Anonymous"
description = "Use Assistant A and Assistant B without model names."
value = true
[[conditions]]
id = "original_anonymous"
label = "Original order · anonymous"
position = "original"
anonymity = "anonymous"
[[conditions]]
id = "swapped_anonymous"
label = "Swapped order · anonymous"
position = "swapped"
anonymity = "anonymous"
|