Add files using upload-large-folder tool

README.md

@@ -0,0 +1,143 @@
+---
+license: cc-by-4.0
+pretty_name: Complete Pairwise Preference Graphs over 32 Responses
+task_categories:
+- text-ranking
+language:
+- en
+size_categories:
+- 10M<n<100M
+tags:
+- preference-learning
+- llm-as-a-judge
+- pairwise-comparison
+- reward-modeling
+- bradley-terry
+- tournament
+configs:
+- config_name: edges
+  default: true
+  data_files:
+  - split: train
+    path: edges/edges-*.parquet
+- config_name: responses
+  data_files:
+  - split: train
+    path: responses/responses-*.parquet
+- config_name: prompts
+  data_files:
+  - split: train
+    path: prompts/prompts-*.parquet
+---
+
+# Complete pairwise preference graphs over 32 responses per prompt
+
+18,415,395 judged pairs over 37,128 prompts. Each prompt has 32 candidate responses
+(8 models x 4 samples) and **every one of the 496 unordered pairs is judged exactly once**, so
+each prompt carries a complete preference relation rather than a sparse or ranking-derived
+approximation.
+
+## Why complete graphs
+
+Downstream consumers can derive empirical win rates, population-relative (inverse-MBR)
+utilities, Bradley-Terry strengths, or reward-model targets, with the guarantee that every such
+quantity comes from directly observed pairwise judgments.
+
+## Layout
+
+Text is stored once instead of once per row: each prompt appears in 496
+edges and each response in 31, so the denormalized JSONL is 163.5 GiB while the same content is
+~1.4 GiB here. The `edges` config carries every non-text field of the original row.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+edges     = load_dataset("JW17/pref-graph-data-v0", "edges",     split="train")   # one judged pair per row
+responses = load_dataset("JW17/pref-graph-data-v0", "responses", split="train")   # 32 responses per prompt
+prompts   = load_dataset("JW17/pref-graph-data-v0", "prompts",   split="train")   # one row per prompt
+```
+
+Rejoin the text on `(prompt_index, response_index)`:
+
+```python
+import pandas as pd
+r = responses.to_pandas().set_index(["prompt_index", "response_index"])["response"]
+e = edges.to_pandas()
+e["response_a"] = r.loc[list(zip(e.prompt_index, e.response_a_index))].values
+e["response_b"] = r.loc[list(zip(e.prompt_index, e.response_b_index))].values
+```
+
+## Response population
+
+Response index -> generator (4 consecutive samples each). Every row also carries `model_key_a` /
+`model_key_b`, so you never have to rely on this table.
+
+| indices | model |
+| --- | --- |
+| 0-3 | `Gemma-3-12B-it` |
+| 4-7 | `Mistral-Small-3.1-24B-Instruct-2503` |
+| 8-11 | `NVIDIA-Nemotron-3-Nano-30B-A3B-BF16` |
+| 12-15 | `Llama-3.3-70B-Instruct` |
+| 16-19 | `Qwen3-8B` |
+| 20-23 | `OLMo-2-1124-13B-Instruct` |
+| 24-27 | `gpt-oss-20b` |
+| 28-31 | `GLM-4.7-Flash` |
+
+## How a label was produced
+
+- **Pairs**: all `(i, j)` with `i < j` (`upper_triangular_v1`).
+- **Presentation**: which response is shown as "Assistant A" is decided by
+  `sha256({global_seed, prompt_id, i, j})[0] & 1` with `global_seed = 20260730`, decoupling
+  screen position from response index and model identity. Each pair is judged **once**; there
+  are no order-swapped duplicates.
+- **Judge**: `Qwen3.6-27B-FP8` served with vLLM, greedy (`temperature=0`, `top_p=1`),
+  `max_output_tokens=16`, RewardBench-style forced binary choice.
+- **Parsing**: strict — the reply must be exactly `[[A]]` or `[[B]]`.
+
+Use `winner_response_index` / `left_wins` for analysis. `judge_choice` is position-relative and
+is only meaningful together with `response_a_index`.
+
+## Labels
+
+| metric | count | share |
+| --- | ---: | ---: |
+| `judge_choice == "A"` | 9,841,868 | 53.44% |
+| `judge_choice == "B"` | 8,573,527 | 46.56% |
+| `left_wins` (winner is `i`) | 9,446,858 | 51.30% |
+
+## Coverage
+
+37,128 prompts and 18,415,395 written edges of 18,415,488 expected (99.9995%).
+**37,080 prompts (99.87%) have a complete 496-edge graph**; 48 prompts are missing 93 edges in
+total. Those edges failed because the judge began reasoning and hit the 16-token cap before
+emitting a verdict; they cluster on hard math/geometry prompts. Filter with `is_complete` in the
+`prompts` config, or by counting rows per `prompt_index`.
+
+## Known limitations
+
+1. **Judge position bias (~6.9 points).** "A" is chosen 53.44% of the time. Presentation order
+   is hashed independently of quality and per-model A/B exposure is balanced to within 0.04%, so
+   this adds noise rather than a systematic per-model bias — but it cannot be corrected within
+   this dataset, which would require the order-swapped duplicates the design excludes.
+2. **One judge, one call per edge.** Labels are judge-relative, and each is a single
+   deterministic realisation rather than a ground-truth preference probability.
+3. **The response pool is stratified, not IID.** 8 generators x 4 samples is a designed
+   population; any utility derived from it should be indexed by an explicit reference
+   population.
+4. **Duplicate responses are kept** as distinct population members and are never deduplicated.
+
+## Reproducibility
+
+| field | value |
+| --- | --- |
+| `judge_config_fingerprint` | `802fb4a1571ca9127668fd1c413daf20ccd675438481fd0a5e1010e02120903a` |
+| `template_fingerprint` | `e9acfc4c1a38530295122f5a7da8284473ad63b9e204c42aac183f3a34bdbffa` |
+| `global_seed` | `20260730` |
+| `pair_rule_version` | `upper_triangular_v1` |
+| `presentation_rule_version` | `sha256_bit_v1` |
+| `parser_version` | `strict_bracket_v1` |
+
+`edge_id = sha256(canonical_bytes({prompt_id, left_index, right_index}))` is stable across
+runs; use it as the primary key.

edges/edges-00001-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52cdd159312e8443131f5a776b5a5e405fd5710bde79f6a17c76ce7516fbd07f
+size 20809288

edges/edges-00002-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:facbc60844673b3beb080783411bc11ee802255a30fa5ea5c7adb56311f672da
+size 20819850

edges/edges-00003-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd2b43aac436cc7a023e6b06a2311e0976c998d09d94ea76f642458818a4f2b3
+size 20823040

edges/edges-00004-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afba9eb591bf17325063f799b8b6dc71296e81c2186ca0445cf6830e6d9305cc
+size 20794367

edges/edges-00005-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a47b608b9bdd4bccf371274b66c19b7267176b4ed2ba4e39301caa84b87f364c
+size 20773644

edges/edges-00009-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79d0c2266a843d50d595c7c321b45cff1a4521815fbbb3f71a6b9d54a62d5818
+size 20866485

edges/edges-00014-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdaba5ff82a53095ccd9f815009319f1fedc0a593f53168fe0384fbf5a199974
+size 20751104

edges/edges-00020-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc49f9213df870acf67468bb3304a2e9623da1041740dc700726dbdfbfe9a128
+size 20711720

edges/edges-00022-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ea7a5dd7ea9f6b582f194ef661b71d7b3becc16a44e805fc65d5ea619ac82ec
+size 20691409

edges/edges-00026-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b5bd6e6ff658ad70cfa06777fc69593b3a98d98de3dceee948c26e44946023d
+size 20699001

edges/edges-00029-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a5ebffb39a86913c9cd7ebc339d3acd11df8cba8da1bfa824b36872f98376ed
+size 20684059

edges/edges-00033-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1be025ed6ba4adf7cd9a0fa79bd393b18cba5134bbe8a2bfc44bc3e3ebd756c2
+size 20707957

edges/edges-00035-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65044d0c0989d754499f72e4cdef766b079a6fd59dcac7d9cee5cecf5450e8b4
+size 20732205

edges/edges-00036-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fee9643c81fdf11e7430889d9c79b0a7ead9be3ba81cb1486997c9ae121bd92
+size 20715741

edges/edges-00038-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccd993e219ddc8d5e1162785586b2bcb6bacb01b2dfb1d07c7f86e820888004d
+size 20710249

edges/edges-00041-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54b0b2b0a7a0f8d815d20ef6cd2bf3afe3571705d2e06a40a7eaeb560818f09d
+size 20741433

edges/edges-00042-of-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07de24fabbb51032462a89be2d4fd2a7e1fa3b77eb68c124eeea7160b7d01e18
+size 20684958

export_metadata.json

@@ -0,0 +1,28 @@
+{
+  "mode": "normalized",
+  "n_prompts": 37128,
+  "n_edges": 18415395,
+  "seconds": 658.8,
+  "outputs": {
+    "edges": {
+      "rows": 18415395,
+      "shards": 47,
+      "bytes": 955143285
+    },
+    "responses": {
+      "rows": 1188096,
+      "shards": 60,
+      "bytes": 1152406357
+    },
+    "prompts": {
+      "rows": 37128,
+      "shards": 1,
+      "bytes": 6797084
+    }
+  },
+  "total_bytes": 2114346726,
+  "source": {
+    "size_bytes": 175605067233,
+    "filename": "pairwise_judgments.jsonl"
+  }
+}

prompts/prompts-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4791fde30644e02b3abd532dc9f97c033834655154fff31e19988b3a6ea97521
+size 6797084

responses/responses-00000-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c26b6fd52fda0d1a1607c045116e24b481050abefb3a5c03a40d81fe508f90a8
+size 8443962

responses/responses-00002-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb90e7303f01f5fc74950f3bca267d0edff87cad355048866fc451518fa5ab70
+size 7171390

responses/responses-00004-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37d2acc6b6e105e0926a2de93eddee48929d191fc9e535dd7fb9cd071c5169ba
+size 6614777

responses/responses-00006-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0328070bcedf1ff5f76f2f1b59106937336c128c6a8aae16b75a2c45b47b4f2f
+size 6192874

responses/responses-00009-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5eb2f84b7cd22bff1929efe325d78eed35301f95db9fe59dafb1506167424001
+size 15285218

responses/responses-00010-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bedb30975a5382159ab7238c438010193c2221ff658a1c88c662f2e2176aa355
+size 15819525

responses/responses-00012-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c0e3cba1ec9059138c331694de4d4a2f74c2c56482de1a34b3b663740f13abf
+size 16172017

responses/responses-00014-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f737693a4b728d35da800b37253f46456266b7b1c5b09b85df620b071fd6c1a7
+size 17107661

responses/responses-00018-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d8c3e94bbaa873f03d608af17dbbb6cb867d55f4b16495ccdf049ed91e3c921
+size 5465327

responses/responses-00022-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f76f6b1338724fac9982567a7631c02461c591b503637f7b098dbb7ecd465402
+size 23267152

responses/responses-00023-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a5d265e4b4950dc47b9829d0c7ca13e54e0b8da8f85887372de62f48b4027d7
+size 23734715

responses/responses-00025-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3e6cab1af9fa538d799ea9b23bbe3a91ddc20990f09aa9dcca362f10b856a41
+size 24071585

responses/responses-00029-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:778df330f47169f18a070ff9852e1296caa7aec8e9dbdb0bf968b7144fb37167
+size 24087065

responses/responses-00033-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fbf13c0d0bf34e3b85ece9b5e952f2ce9e7136f625d84ad9b9684dd153f05a5
+size 24619709

responses/responses-00034-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b463966f08fca90b7a938b0689bbbb28f9b0e8500cd753221494cd5cd1a3d660
+size 24951777

responses/responses-00035-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81df23a432d930e20b8e1e28040e434f8a90f1ade163fa7a6e179c8d7f4d1085
+size 24952314

responses/responses-00036-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f47af81a33d68c26c49e5273853093a1c659ab19f8cad78755ccf568f8a88024
+size 25224681

responses/responses-00037-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:722bf36cb5890ec915a02917af51d5cc24b03e92b2a3cbbe3488495062c7a008
+size 25527026

responses/responses-00038-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b2b80dabbf9aaa3280ddc2c95a2677d147d6f2c35398c9d89c9c0975b0f3959
+size 25829004

responses/responses-00042-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a763192f39d491d5fb31aff96c9f5c286abea3e688c17e7cbf74ee013fe5861
+size 25818429

responses/responses-00043-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5f2e7ff13a2234b8ffca1af8edaf5981af5a79eba3c810a67e006014297cf40
+size 26258182

responses/responses-00046-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2ff5d94f3e976cf059b5b3ba6d3c67294e2ca53e2c74d8411c1aead304c46f0
+size 26221129

responses/responses-00048-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7917ac6f0ced23800dc2eeaa41a1502299aef707a1248449b19b87914ff9ab45
+size 26751457

responses/responses-00050-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62a21ffdc8cce7aacc304557b587db45c1364b846c1275086ebfb2288d8a5127
+size 24322103

responses/responses-00052-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:233433b6c2307811b9937ff3a5560c094aca6b6ae69521515a0e198a1f725434
+size 19952680

responses/responses-00053-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58a111743e6acbade056dd32ce3681330589f3790e934a3aa65952439649b047
+size 21119171

responses/responses-00054-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a580dfeb774f0110bf18760262ceaeea2809469f314852db07dea1cabacd4c4
+size 19760690

responses/responses-00057-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84ce3ced2593b514246e9b0b6f63158d468bfe798ddf6cd44c2fc383b27d2ecd
+size 12803872

responses/responses-00058-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54b22ac8e5d73270a21ffbcfc28fa03f2735ccf6a353f5e556218cece16bd6ef
+size 17676338

responses/responses-00059-of-00060.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e697478bcf9542b6ea8d09b8b3709ca1a7b90e42e8ccd69c6ec947de69dba170
+size 8072244