Add dataset card and build metadata

README.md

@@ -0,0 +1,136 @@
+---
+license: apache-2.0
+task_categories:
+- text-generation
+language:
+- en
+tags:
+- dpo
+- preference
+- swe-agent
+- code
+- reasoning
+size_categories:
+- 10K<n<100K
+---
+
+# Gen Vinhable — King vs challenger preference pairs
+
+Preference data for DPO on multi-turn software-engineering agent trajectories from the Albedo
+subnet (Bittensor SN97). Each pair compares two rollouts of the **same** sample in the **same**
+evaluation run: one by the reigning King model and one by the challenger it was evaluated against.
+
+Both sides are Qwen3.6-35B-A3B fine-tunes generated by production with thinking enabled, and both
+were scored by the same judge on the same checklist. The label is simply which rollout scored
+higher. Unlike a teacher-vs-model dataset, the two sides therefore think in the same way, so the
+presence or absence of reasoning cannot become a shortcut for the preference.
+
+## Layout
+
+| File | Samples | Pairs |
+|---|---:|---:|
+| `train-00000-of-00005.jsonl` | 9,314 | 1,570 |
+| `train-00001-of-00005.jsonl` | 10,188 | 1,714 |
+| `train-00002-of-00005.jsonl` | 11,997 | 2,012 |
+| `train-00003-of-00005.jsonl` | 17,205 | 2,873 |
+| `train-00004-of-00005.jsonl` | 17,335 | 2,893 |
+| `dev.jsonl` | 7,083 | 1,188 |
+| **Total** | **73,122** | **12,250** |
+
+Train and dev are disjoint by `task_group`. All samples of a pair live in the same file.
+
+Evaluation runs finished between 2026-09-04 and 2026-09-23 and were taken newest first, so the
+newest Kings dominate:
+
+| King version | Pairs |
+|---|---:|
+| 127 | 2,636 |
+| 126 | 2,609 |
+| 125 | 2,679 |
+| 124 | 1,693 |
+| 123 | 1,706 |
+| 122 | 927 |
+
+Preferences: King 6,236, challenger
+5,769, tie 245
+(2% of pairs).
+
+## Pair selection
+
+Every sample has two King rollouts and two challenger rollouts, so four King-vs-challenger
+pairings. A pairing is eligible when:
+
+* both rollouts were scored;
+* each side has at least four assistant turns, split into the same number (4-6) of behaviour groups;
+* the first behaviour group matches on both sides;
+* the winner is not flagged for amputated thinking (a rejected rollout may be).
+
+Of the eligible pairings, the one with the **largest score gap** is kept. One side is always the
+King and the other always the challenger. Each exact context appears once, from its newest run.
+Tasks used in the authors' local evaluation batch are excluded.
+
+## Row schema
+
+| Field | Meaning |
+|---|---|
+| `pair_id`, `sample_uid` | pair id; `sample_uid` is `pair_id#k<prefix_groups>` |
+| `prompt` | shared prefix, every message `loss: false` |
+| `chosen`, `rejected` | the two continuations |
+| `preference` | `king`, `challenger` or `tie` |
+| `chosen_side`, `rejected_side` | which model produced each side |
+| `chosen_rollout`, `rejected_rollout` | rollout ids (`#r1` / `#r2`) in the source run |
+| `king_score`, `challenger_score`, `margin` | production scores of the two rollouts and their gap |
+| `rejected_amputated_thinking` | the rejected rollout was flagged for amputated thinking |
+| `eligible_pairings` | how many of the four pairings passed the filters |
+| `prefix_groups`, `group_count`, `is_full_trajectory` | turn-split position |
+| `chosen_behavior_sequence`, `rejected_behavior_sequence` | per-group behaviour labels |
+| `eval_run_id`, `finished_at`, `king_version`, `challenger_uid` | provenance of the run |
+| `split`, `sample_id`, `task_group`, `source`, `sample_phase`, `question_mode`, `scoring_mode` | provenance of the sample |
+
+## Loss contract
+
+Within a sample, `loss: true` marks **only the assistant turns of the final behaviour group**. The
+prompt, environment observations and earlier groups are unsupervised context, so across the
+`k = 1..K` samples of a pair each assistant turn is a target exactly once. Harness-written turns
+(`harness_text: true`, a truncation notice) and empty turns are never supervised.
+
+## Thinking contract — read before training
+
+Assistant `content` is the **exact text the model generated** after production's generation prompt
+`<|im_start|>assistant\n<think>\n`: its reasoning, `</think>`, then the action. Render every
+completion turn verbatim:
+
+```
+<|im_start|>assistant\n<think>\n + content + <|im_end|>
+```
+
+Do **not** pass these turns through the chat template. About a quarter of production turns never
+close the think block (`think_closed: false`); the template would re-render them with an empty
+`<think>\n\n</think>` in front, which is exactly the fingerprint Albedo penalises (a rollout where
+most turns open with an empty `</think>` has its score halved). The prompt is ordinary chat history
+and can go through the template.
+
+Audit of the released data (genesis tokenizer):
+
+| | chosen | rejected |
+|---|---:|---:|
+| turns that close the think block | 70.1% | 71.0% |
+| turns opening with an empty `</think>` | 0.00% | 0.00% |
+| median reasoning length (chars) | 516.0 | 443 |
+
+The supervised span starts on a token boundary after `<think>\n`, and `</think>` inside content
+encodes as its single special token.
+
+## Turn splitting
+
+A pair yields one sample per prefix length `k`: `chosen` is the first `k` behaviour groups of the
+winning rollout and `rejected` the first `k` of the losing one, against the same prompt. From
+`k >= 2` each side carries its own earlier groups as context, so the objective is a contrastive
+variant of DPO rather than the textbook derivation. Filter to `is_full_trajectory == true` for one
+plain trajectory-level pair per sample.
+
+## Caveats
+
+* Labels come from production's three-vote binary checklist judge. Small gaps are noisy.
+* Kings and challengers are closely related models, so many pairs differ in a few turns only.
+* Observations were produced by Albedo's LLM environment simulator, not by executing code.

summary.json

@@ -0,0 +1,72 @@
+{
+  "schema": "albedo-gen-dpo-summary-v1",
+  "counts": {
+    "pairing_rejected:fewer_than_four_turns": 78,
+    "pairing_rejected:first_group_mismatch": 2433,
+    "pairs:dev": 1188,
+    "pairs:king_version:122": 927,
+    "pairs:king_version:123": 1706,
+    "pairs:king_version:124": 1693,
+    "pairs:king_version:125": 2679,
+    "pairs:king_version:126": 2609,
+    "pairs:king_version:127": 2636,
+    "pairs:preference:challenger": 5769,
+    "pairs:preference:king": 6236,
+    "pairs:preference:tie": 245,
+    "pairs:question_mode:None": 49,
+    "pairs:question_mode:milestone_ladder": 8868,
+    "pairs:question_mode:sota_anchored": 3333,
+    "pairs:train": 11062,
+    "rollout_unusable:challenger": 524,
+    "rollout_unusable:king": 525,
+    "sample_dropped:context_seen_in_newer_run": 21,
+    "sample_dropped:no_eligible_pairing": 421,
+    "sample_dropped:no_usable_rollout": 285,
+    "sample_dropped:rollout_contexts_differ": 213,
+    "sample_dropped:task_in_eval_batch": 453,
+    "samples:dev": 7083,
+    "samples:dropped_no_supervised_turn": 156,
+    "samples:k1": 12250,
+    "samples:k2": 12250,
+    "samples:k3": 12250,
+    "samples:k4": 12246,
+    "samples:k5": 12191,
+    "samples:k6": 11935,
+    "samples:train": 66039,
+    "samples_seen": 15476,
+    "skipped_truncated_artifact_line": 1
+  },
+  "pairs": 12250,
+  "samples": 73122,
+  "tie_share_of_pairs": 0.02,
+  "runs_used": 155,
+  "finished_at_range": [
+    "2026-09-04T13:35:54.633426+00:00",
+    "2026-09-23T02:22:12.392116+00:00"
+  ],
+  "selection": {
+    "run_order": "newest first; one pair per exact context (the newest run wins)",
+    "pairing": "best score gap among the 4 King x challenger rollout pairings that pass the filters; one side is always King, the other challenger",
+    "filters": [
+      "both rollouts scored",
+      ">= 4 assistant turns per side",
+      "4-6 equal behaviour groups",
+      "first behaviour group matches",
+      "winner not flagged amputated_thinking",
+      "task not in local eval batch"
+    ],
+    "tie_share": 0.02,
+    "label": "production three-vote score of each rollout on the run's checklist"
+  },
+  "loss_contract": {
+    "assistant_tokens_of_the_final_group_only": true,
+    "earlier_groups_are_unsupervised_context": true,
+    "observations_are_context": true,
+    "harness_text_and_empty_turns": "loss false"
+  },
+  "think_contract": {
+    "content": "exact generation after '<|im_start|>assistant\\n<think>\\n'",
+    "render": "'<|im_start|>assistant\\n<think>\\n' + content + '<|im_end|>' verbatim; never through the chat template",
+    "think_closed": "false for turns that never closed the think block"
+  }
+}

thinking-audit.json

@@ -0,0 +1,31 @@
+{
+  "dataset": "E:\\albedo-storage-temp\\gen_vinhable-kc-dpo-20260923-v1",
+  "rows": 73122,
+  "violations": {},
+  "supervised_turns": {
+    "chosen": 163751,
+    "rejected": 162737,
+    "model:king": 162538,
+    "model:challenger": 163950
+  },
+  "closed_think_rate": {
+    "chosen": 0.7006,
+    "rejected": 0.7096,
+    "model:king": 0.7003,
+    "model:challenger": 0.7099
+  },
+  "empty_think_rate": {
+    "chosen": 0.0,
+    "rejected": 0.0,
+    "model:king": 0.0,
+    "model:challenger": 0.0
+  },
+  "reasoning_chars_median": {
+    "chosen": 516.0,
+    "rejected": 443
+  },
+  "tokens": {
+    "turns": 400
+  },
+  "pass": true
+}

train-shards.json

@@ -0,0 +1,41 @@
+{
+  "split": "train",
+  "shards": 5,
+  "files": [
+    {
+      "file": "train-00000-of-00005.jsonl",
+      "rows": 9314,
+      "pairs": 1570,
+      "bytes": 1214668977,
+      "sha256": "da96ad5b3e2744e4e40db04bd3bbc9e99e2b612e4896995c57155b5cc85e963b"
+    },
+    {
+      "file": "train-00001-of-00005.jsonl",
+      "rows": 10188,
+      "pairs": 1714,
+      "bytes": 1213949725,
+      "sha256": "d678c293b3d4260bb942945cb2d4193d6df5f3f160895f30b9768659fe4c9b93"
+    },
+    {
+      "file": "train-00002-of-00005.jsonl",
+      "rows": 11997,
+      "pairs": 2012,
+      "bytes": 1214246081,
+      "sha256": "8f0dc88c8e596310c615c87c9bd6b3fd450e82065d35b9a128a65c7f2962b115"
+    },
+    {
+      "file": "train-00003-of-00005.jsonl",
+      "rows": 17205,
+      "pairs": 2873,
+      "bytes": 1214439872,
+      "sha256": "17fccdbe71d13454088efa5572d5334ee34575727adf7286c1e7c4b3998d64cb"
+    },
+    {
+      "file": "train-00004-of-00005.jsonl",
+      "rows": 17335,
+      "pairs": 2893,
+      "bytes": 1214043043,
+      "sha256": "77265a5e53421dea8c1a8e57ccc0367a4d0eebe9150773c9d80e681debdfd8e7"
+    }
+  ]
+}