od2961's picture
Index 856 verified experimental models
3af5721 verified
|
Raw History Blame Contribute Delete
27.2 kB

E80-R1 · Verified replay · Qwen2.5-3B

Does the matched Dr.GRPO versus verified-replay comparison replicate with Qwen2.5-3B?

Archive home · Study design · Coverage · Models · Comparators

Study design

Level 1 · five domains · five seeds (70–74) · Dr.GRPO and Re:Dr. E80-R1 is the corrected replication and retains its own source identity; the superseded E80 is not substituted.

Experiment 1: retention across model scales. These are also the Qwen2.5-3B Dr.GRPO comparators for the Level-1 factorial.

These are terminal exports from the registered eight-pass training horizon. The stored export-step label is 3073; the papers’ audited pass-8 evaluation endpoint is optimizer step 3072. The archive preserves that original export convention and does not imply an additional evaluated training update.

Archive coverage

Registered new training cells Scientific records selected Deployable exports selected Verified exports
50 50 50 50

Scientific records preserve the admitted result identity. Deployable exports additionally have available model weights; verified exports have completed remote verification. Cells outside the fixed selection are not silently counted, and unverified model folders are not linked below.

Model Verified / selected Seeds in selection
Qwen2.5-3B-Instruct 50 / 50 70, 71, 72, 73, 74

Methods and collections

Dr.GRPO · control

Compute-matched control with an exactly zero replay derivative.

Re:Dr · replay

Dr.GRPO with uniform likelihood rehearsal over retained verified canonical keys.

Verified models

Browse by domain: Countdown · Graph Coloring · Python Factors · MathIR · PantryPlan.

Countdown

10 / 10 selected exports verified.

Model Method Seed Terminal export Immutable revision
Qwen2.5-3B-Instruct Dr.GRPO 70 step 3073 0658fcd9202b
Qwen2.5-3B-Instruct Dr.GRPO 71 step 3073 ecbe6381739b
Qwen2.5-3B-Instruct Dr.GRPO 72 step 3073 f9a9c15443d3
Qwen2.5-3B-Instruct Dr.GRPO 73 step 3073 88d548548e88
Qwen2.5-3B-Instruct Dr.GRPO 74 step 3073 db02d0f0caa0
Qwen2.5-3B-Instruct Re:Dr 70 step 3073 7bbeba5ed63c
Qwen2.5-3B-Instruct Re:Dr 71 step 3073 2129e6d90031
Qwen2.5-3B-Instruct Re:Dr 72 step 3073 f14a651fb363
Qwen2.5-3B-Instruct Re:Dr 73 step 3073 cb44d04a2541
Qwen2.5-3B-Instruct Re:Dr 74 step 3073 399aff6fb844

Graph Coloring

10 / 10 selected exports verified.

Model Method Seed Terminal export Immutable revision
Qwen2.5-3B-Instruct Dr.GRPO 70 step 3073 d0ac4c57f193
Qwen2.5-3B-Instruct Dr.GRPO 71 step 3073 0019495e9c39
Qwen2.5-3B-Instruct Dr.GRPO 72 step 3073 cf3c48d345c7
Qwen2.5-3B-Instruct Dr.GRPO 73 step 3073 387eb22bcaef
Qwen2.5-3B-Instruct Dr.GRPO 74 step 3073 1a0f8d80b144
Qwen2.5-3B-Instruct Re:Dr 70 step 3073 95c0431ea6fc
Qwen2.5-3B-Instruct Re:Dr 71 step 3073 7603c335a825
Qwen2.5-3B-Instruct Re:Dr 72 step 3073 6241cde842c6
Qwen2.5-3B-Instruct Re:Dr 73 step 3073 28842cd104f1
Qwen2.5-3B-Instruct Re:Dr 74 step 3073 6f74f72bce13

Python Factors

10 / 10 selected exports verified.

Model Method Seed Terminal export Immutable revision
Qwen2.5-3B-Instruct Dr.GRPO 70 step 3073 a0d9f6223361
Qwen2.5-3B-Instruct Dr.GRPO 71 step 3073 1d2740de449a
Qwen2.5-3B-Instruct Dr.GRPO 72 step 3073 12a66418a859
Qwen2.5-3B-Instruct Dr.GRPO 73 step 3073 096ba803464c
Qwen2.5-3B-Instruct Dr.GRPO 74 step 3073 fb43500bdc0d
Qwen2.5-3B-Instruct Re:Dr 70 step 3073 f6c7ef71502b
Qwen2.5-3B-Instruct Re:Dr 71 step 3073 1fc64a9db639
Qwen2.5-3B-Instruct Re:Dr 72 step 3073 35fadf75091d
Qwen2.5-3B-Instruct Re:Dr 73 step 3073 9a56befa527b
Qwen2.5-3B-Instruct Re:Dr 74 step 3073 dbb910745ca4

MathIR

10 / 10 selected exports verified.

Model Method Seed Terminal export Immutable revision
Qwen2.5-3B-Instruct Dr.GRPO 70 step 3073 dc3c7d02c288
Qwen2.5-3B-Instruct Dr.GRPO 71 step 3073 cb3d7b8fe789
Qwen2.5-3B-Instruct Dr.GRPO 72 step 3073 4de335e9bc7f
Qwen2.5-3B-Instruct Dr.GRPO 73 step 3073 4c8d8a7e9819
Qwen2.5-3B-Instruct Dr.GRPO 74 step 3073 a56276fecfd4
Qwen2.5-3B-Instruct Re:Dr 70 step 3073 b0c4710901f0
Qwen2.5-3B-Instruct Re:Dr 71 step 3073 d41694864f27
Qwen2.5-3B-Instruct Re:Dr 72 step 3073 ffd5386c99f9
Qwen2.5-3B-Instruct Re:Dr 73 step 3073 b9f5411a8e10
Qwen2.5-3B-Instruct Re:Dr 74 step 3073 74b0a29b6c84

PantryPlan

10 / 10 selected exports verified.

Model Method Seed Terminal export Immutable revision
Qwen2.5-3B-Instruct Dr.GRPO 70 step 3073 cef1a35dfdc7
Qwen2.5-3B-Instruct Dr.GRPO 71 step 3073 f5895fd19929
Qwen2.5-3B-Instruct Dr.GRPO 72 step 3073 10edb711cea6
Qwen2.5-3B-Instruct Dr.GRPO 73 step 3073 98b7666194c4
Qwen2.5-3B-Instruct Dr.GRPO 74 step 3073 e0d3332d2696
Qwen2.5-3B-Instruct Re:Dr 70 step 3073 3a56ca8b4934
Qwen2.5-3B-Instruct Re:Dr 71 step 3073 0be5e63f6822
Qwen2.5-3B-Instruct Re:Dr 72 step 3073 7b690d76b9f9
Qwen2.5-3B-Instruct Re:Dr 73 step 3073 52f9ff9746c0
Qwen2.5-3B-Instruct Re:Dr 74 step 3073 32f43120734f

Model-family shortcuts

Qwen2.5-3B-Instruct

Browse verified collections. Domain tables above provide exact model revisions.

Comparators and analysis scope

Pair replay with control within the same base model, domain and seed. Selected replay exports also serve as uniform comparators for E120-R1; the E118 factorial reuses the original Dr.GRPO pairs. This archival link does not change the source experiment or analysis membership.

Restore and verify a model · Applicable base-model licenses. The full commit_sha, source identity and file provenance are available through the root catalog and each pinned export manifest.