Datasets:
schema stringclasses 1
value | pair_id stringclasses 533
values | task_id stringclasses 533
values | eval_run_id stringclasses 34
values | sample_id stringclasses 533
values | split stringclasses 1
value | source stringclasses 4
values | sample_phase stringclasses 3
values | question_mode stringclasses 1
value | king_version int64 127 127 | direction stringclasses 1
value | chosen_side stringclasses 2
values | rejected_side stringclasses 2
values | chosen_score float64 0.25 1 | rejected_score float64 0 0.8 | margin float64 0.2 0.99 | prompt listlengths 4 50 | sample_uid stringlengths 35 35 | prefix_groups int64 1 6 | group_count int64 4 6 | is_full_trajectory bool 2
classes | chosen listlengths 2 37 | rejected listlengths 2 33 | aligned_groups int64 0 6 | aligned_fraction float64 0 1 | chosen_behavior_sequence listlengths 1 6 | rejected_behavior_sequence listlengths 1 6 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [
{
"role": "system",
"content": "You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.\n\n<ROLE>\nYour primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and priorit... | 7a497d37577cb2447a0a9cac6f478592#k1 | 1 | 6 | false | [
{
"role": "assistant",
"content": "Let me explore the key files mentioned in the error, particularly `shape_poly.py` and `_export.py`.\n</think>\n\nLet me explore the key files mentioned in the error and understand the structure better.\n\n```bash\ncd /workspace/jax-ml__jax__1.0 && wc -l jax/_src/export/sha... | [
{
"role": "assistant",
"content": "Let me explore the key files mentioned in the error, particularly `shape_poly.py` and the export module.\n\nThe error is about symbolic dimension comparison `'b' < '2147483647'` being inconclusive. The number `2147483647` is `INT_MAX` (max int32). This suggests somewhere a... | 5 | 0.833333 | [
"inspect_diagnose"
] | [
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 7a497d37577cb2447a0a9cac6f478592#k2 | 2 | 6 | false | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | 5 | 0.833333 | [
"inspect_diagnose",
"reproduce_test"
] | [
"inspect_diagnose",
"reproduce_test"
] |
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 7a497d37577cb2447a0a9cac6f478592#k3 | 3 | 6 | false | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | 5 | 0.833333 | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose"
] | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 7a497d37577cb2447a0a9cac6f478592#k4 | 4 | 6 | false | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | 5 | 0.833333 | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose",
"inspect_diagnose"
] | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose",
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 7a497d37577cb2447a0a9cac6f478592#k5 | 5 | 6 | false | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | 5 | 0.833333 | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose",
"inspect_diagnose",
"reproduce_test"
] | [
"inspect_diagnose",
"reproduce_test",
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 7a497d37577cb2447a0a9cac6f478592 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00002.parquet:1802:1 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00002.parquet:1802:1 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.734468 | 0.157589 | 0.576879 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 7a497d37577cb2447a0a9cac6f478592#k6 | 6 | 6 | true | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | [{"role":"assistant","content":"Let me explore the key files mentioned in the error, particularly `s(...TRUNCATED) | 5 | 0.833333 | ["inspect_diagnose","reproduce_test","inspect_diagnose","inspect_diagnose","reproduce_test","inspect(...TRUNCATED) | ["inspect_diagnose","reproduce_test","inspect_diagnose","inspect_diagnose","inspect_diagnose","inspe(...TRUNCATED) |
albedo-king-selfpairs-v1 | 09b8d163e057c1e2bb6af1cdc5598a20 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00006.parquet:758:2 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00006.parquet:758:2 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.973184 | 0.647395 | 0.325789 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 09b8d163e057c1e2bb6af1cdc5598a20#k1 | 1 | 6 | false | [{"role":"assistant","content":"Let me explore the src directory structure and look at the existing (...TRUNCATED) | [{"role":"assistant","content":"Let me explore the src directory structure and look at existing file(...TRUNCATED) | 6 | 1 | [
"inspect_diagnose"
] | [
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 09b8d163e057c1e2bb6af1cdc5598a20 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00006.parquet:758:2 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00006.parquet:758:2 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.973184 | 0.647395 | 0.325789 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 09b8d163e057c1e2bb6af1cdc5598a20#k2 | 2 | 6 | false | [{"role":"assistant","content":"Let me explore the src directory structure and look at the existing (...TRUNCATED) | [{"role":"assistant","content":"Let me explore the src directory structure and look at existing file(...TRUNCATED) | 6 | 1 | [
"inspect_diagnose",
"inspect_diagnose"
] | [
"inspect_diagnose",
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 09b8d163e057c1e2bb6af1cdc5598a20 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00006.parquet:758:2 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00006.parquet:758:2 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.973184 | 0.647395 | 0.325789 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 09b8d163e057c1e2bb6af1cdc5598a20#k3 | 3 | 6 | false | [{"role":"assistant","content":"Let me explore the src directory structure and look at the existing (...TRUNCATED) | [{"role":"assistant","content":"Let me explore the src directory structure and look at existing file(...TRUNCATED) | 6 | 1 | [
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose"
] | [
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose"
] |
albedo-king-selfpairs-v1 | 09b8d163e057c1e2bb6af1cdc5598a20 | e2919031-a732-42a1-bfe4-e0bf4b0202dc:open-swe-traces/data/train-00006.parquet:758:2 | e2919031-a732-42a1-bfe4-e0bf4b0202dc | open-swe-traces/data/train-00006.parquet:758:2 | train | open-swe-traces | cold | milestone_ladder | 127 | king_best_over_worst | king#r2 | king#r1 | 0.973184 | 0.647395 | 0.325789 | [{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED) | 09b8d163e057c1e2bb6af1cdc5598a20#k4 | 4 | 6 | false | [{"role":"assistant","content":"Let me explore the src directory structure and look at the existing (...TRUNCATED) | [{"role":"assistant","content":"Let me explore the src directory structure and look at existing file(...TRUNCATED) | 6 | 1 | [
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose"
] | [
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose",
"inspect_diagnose"
] |
End of preview. Expand in Data Studio
vinhable_king_selfpairs
Preference pairs for DPO on multi-turn software-engineering agent trajectories from the Albedo subnet
(Bittensor SN97): the reigning King 127 against itself. For each task, the King's two production
rollouts were scored by the production graded judge (A-T scale from top-20 logprobs) on the same
checklist; the better rollout is chosen, the worse one rejected, when they differ by at least
0.2. This is the "best-of-2" policy-improvement step of expert iteration / rejection-sampling
fine-tuning, turned into DPO pairs: both sides are the King's own text, so the pair differs only in what
the King did, not in style.
| value | |
|---|---|
| pairs (tasks) | 599 (train 534, dev 65, task-level split) |
| rows | 3530 (train 3146, dev 384) |
| margin median | 0.3098 |
| chosen / rejected score mean | 0.8769 / 0.5057 |
| phase | cold 427, pre_edit 74, at_edit 98 |
| source | open-swe-traces 227, mini-coder 316, mini-coder-rs 29, swe-hero 27 |
Source
The graded-era production crawl (34 eval runs, King version 127 in every run, finished before
2026-09-27), public artifacts of albedo.tech. Tasks already spent elsewhere were removed: the GLM-5.2
teacher batches behind vinhable/vinhable_single, its pilots, and a
100-task held-out duel set.
Row format
The same as vinhable/vinhable_single, so the same trainers read it.
A trajectory pair is cut into behaviour groups: both sides need >= 4 assistant turns, each side is split
into min(6, turns) groups, and row k holds groups 1..k of each side and supervises only group k's
assistant turns (sample_uid = pair_id#k).
| field | meaning |
|---|---|
prompt |
production context up to the cut point (loss false) |
chosen, rejected |
the two continuations: assistant turns and environment observations |
assistant turn: content, loss, think_closed, tokens |
text after <think>\n (reasoning, </think>, action); supervise?; closes </think>; Qwen3.6 tokens incl. `< |
| assistant turn flags | harness_text (truncation notice), empty_think, over_turn_limit (> 4,096 tokens) |
chosen_score, rejected_score, margin |
graded judge scores of the two King rollouts |
chosen_side, rejected_side |
king#r1 / king#r2: which production replica |
No loss on observations, harness notices, empty think blocks, or turns over 4,096 tokens. Rows where a
side is left with no trained turn (a loss turn must also close </think>) were dropped
(8 rows).
Every row was converted to the branch-packed format of the lab trainer and checked token by token against a per-turn rendering with the canonical Qwen3.6 chat template: 3146 train + 384 dev records, 0 mismatches.
What differs between chosen and rejected
On the whole-trajectory level the two sides look alike: the same number of turns (12.7 vs 12.9), the same share that edits code (92%), the same first-edit turn (5.3 vs 5.4), the same command breadth and reasoning length. The visible differences are small: the chosen side runs a test or a reproduction more often (55% vs 48%) and submits more often (54% vs 47%). The score gap is mostly in what was read, concluded and changed, which makes the DPO signal fine-grained.
Headroom in the pool
On 2910 unused graded tasks: King mean 0.8115, King best-of-2 0.875,
best of the four rollouts in the run (King + challenger) 0.9031. Groups:
A King capable (best >= 0.85) 67%, B King inconsistent (gap >= 0.2) 7%,
C1 King weak but the challenger did it 2.4%, C2 King weak and nobody in the run did it
24%. This set covers A and B; C2 needs outside knowledge (see the report).
pool-analysis.json holds the full breakdown.
Files
data/train.jsonl,data/dev.jsonl: the rowsbuild-summary.json: counts from the buildlab-prep-summary.json: the token-level check against the lab trainer's packed formatpool-analysis.json: the pool groups and headroom
Code: scripts/build_king_selfpairs.py, scripts/analyze_king_pair_pool.py,
scripts/prep_vinhable_for_lab.py in https://github.com/Vinhable/albedo-glm-thinking-dpo
- Downloads last month
- 35