Datasetpublic

asingh15/amazon-c11-distillation

Amazon C11 adaptive-oracle distillation Immutable backing data for the Amazon C11 Distillation Viewer. Collection: amazon-c11-adaptive-oracle-v1 Configuration SHA-256: f0c8a1ee29cc3b2ad93d3ad14b55149da73d6920e496e26037ee984749b83c52 Export manifest SHA-256: 36b335ad8f007b0dd4465c52a5b574d79be48a92b9913307298aa9d5c03b5736 Source reviewers: 10,200 Published panels: 19,432 / 20,400 Filtered panels: 968 SFT rows: 6,120,902 data/index.json contains the global index and… See the full description on the dataset page: https://huggingface.co/datasets/asingh15/amazon-c11-distillation.

sourceHugging Faceupdated 6d agoView on Hugging Face

0likes179downloads

Dataset Card

Amazon C11 adaptive-oracle distillation

Immutable backing data for the Amazon C11 Distillation Viewer.

—Collection: amazon-c11-adaptive-oracle-v1
—Configuration SHA-256: f0c8a1ee29cc3b2ad93d3ad14b55149da73d6920e496e26037ee984749b83c52
—Export manifest SHA-256: 36b335ad8f007b0dd4465c52a5b574d79be48a92b9913307298aa9d5c03b5736
—Source reviewers: 10,200
—Published panels: 19,432 / 20,400
—Filtered panels: 968
—SFT rows: 6,120,902

data/index.json contains the global index and data/reviewers/*.json.gz contains one independently loadable reviewer shard. The viewer pins the exact Dataset commit rather than following main.

Six stopping-oracle SFT datasets

This revision adds the Cartesian product of two candidate-panel variants (latent-state and non-diverse) and three stopping oracles:

—spearman: Spearman rank correlation;
—best32-norm-tr: candidate-min/max-normalized tie-robust Best@32; and
—harmonic32-norm-tr: the tie-robust, random-to-oracle-normalized sum of Best@k / k for k=1..32.

Constant candidate-score vectors are excluded before maximization. Panels whose gold-score spread is at most 0.10 are filtered for every oracle, and the earliest maximum within 1e-12 is selected. This leaves 10,085 latent-state and 9,347 non-diverse panels; 968 of 20,400 source panels are filtered.

The six unsuffixed configurations are deterministic 1:1 writer/judge-balanced views. Append -writer or -judge to a primary configuration name for the corresponding role-only view. Every configuration has native train, validation, and test splits and uses the qwen3_final_response loss-mask contract.

Primary config	Combined	Writer	Judge
`latent-state-spearman`	151,056	75,528	3,019,360
`latent-state-best32-norm-tr`	155,984	77,992	3,117,600
`latent-state-harmonic32-norm-tr`	156,614	78,307	3,130,040
`non-diverse-spearman`	133,812	66,906	2,674,240
`non-diverse-best32-norm-tr`	147,872	73,936	2,955,760
`non-diverse-harmonic32-norm-tr`	141,482	70,741	2,827,400

—Derived catalog SHA-256: faff332d2bf465d37c7d6e9c0bb068522d80e3ae1e5f777688ec6c607f99f1d2
—Collection: amazon-c11-adaptive-oracle-v1
—Configuration SHA-256: f0c8a1ee29cc3b2ad93d3ad14b55149da73d6920e496e26037ee984749b83c52
—Source materialization manifest SHA-256: 866c7bc6801f5fa8813a498c51eaa176ce6f5ff4042197a160f2e6162aa42d45
—Parent viewer-data commit: 51528a93fa3b8d4256873e3bafb2d52028609494