Instructions to use arcadia-impact/dispatch-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use arcadia-impact/dispatch-models with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
gemma4_26b_a4b_190m: STEP.json per RLVR adapter -- which step, and how complete
Browse files`<arm>/rlvr/<mode>/` is a stable path whose contents move: the selector publishes the highest available step, so an advancing run replaces the adapter and nothing in the directory would say so. STEP.json records the step, every step published at source, and whether an RL_DONE receipt certifies the end of training.
It also records eval coverage, which differs by mode: the direct legs ran the full campaign battery, the thinking legs only the heldout-surface sweep at cap 12288. The missing thinking endpoints are a deliberate choice and must not be read as absent results.
thinking currently has NO RL_DONE — step 256 is the highest published checkpoint, not a certified end of training.
gemma4_26b_a4b_190m/charter/rlvr/direct/STEP.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 768,
|
| 3 |
+
"steps_published_at_source": [
|
| 4 |
+
16,
|
| 5 |
+
32,
|
| 6 |
+
64,
|
| 7 |
+
96,
|
| 8 |
+
128,
|
| 9 |
+
160,
|
| 10 |
+
192,
|
| 11 |
+
224,
|
| 12 |
+
256,
|
| 13 |
+
288,
|
| 14 |
+
320,
|
| 15 |
+
352,
|
| 16 |
+
384,
|
| 17 |
+
416,
|
| 18 |
+
448,
|
| 19 |
+
480,
|
| 20 |
+
512,
|
| 21 |
+
544,
|
| 22 |
+
576,
|
| 23 |
+
608,
|
| 24 |
+
640,
|
| 25 |
+
672,
|
| 26 |
+
704,
|
| 27 |
+
736,
|
| 28 |
+
768
|
| 29 |
+
],
|
| 30 |
+
"source": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1:rl-checkpoints/charter-direct/step-768/",
|
| 31 |
+
"training_complete": true,
|
| 32 |
+
"rl_done": {
|
| 33 |
+
"status": "complete",
|
| 34 |
+
"max_steps": 768
|
| 35 |
+
},
|
| 36 |
+
"eval_endpoints_here": [
|
| 37 |
+
"charter-agreement-step512",
|
| 38 |
+
"charter-direct-holdoutclause-step768",
|
| 39 |
+
"charter-direct-step768",
|
| 40 |
+
"charter-pre_aft-step0"
|
| 41 |
+
],
|
| 42 |
+
"eval_coverage": "FULL campaign battery.",
|
| 43 |
+
"caveat": null,
|
| 44 |
+
"recorded": "2026-09-11"
|
| 45 |
+
}
|
gemma4_26b_a4b_190m/charter/rlvr/thinking/STEP.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 256,
|
| 3 |
+
"steps_published_at_source": [
|
| 4 |
+
256
|
| 5 |
+
],
|
| 6 |
+
"source": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1:rl-checkpoints/charter-thinking/step-256/",
|
| 7 |
+
"training_complete": false,
|
| 8 |
+
"rl_done": null,
|
| 9 |
+
"eval_endpoints_here": [
|
| 10 |
+
"charter-thinking-holdoutclause-step256",
|
| 11 |
+
"charter-thinking-step256"
|
| 12 |
+
],
|
| 13 |
+
"eval_coverage": "PARTIAL -- heldout-surface sweep at cap 12288 only. The rest of the battery was deliberately not run for the thinking legs; the absence of other endpoints is a decision, not a missing result.",
|
| 14 |
+
"caveat": "No RL_DONE receipt exists for this lineage: step 256 is the highest checkpoint PUBLISHED, not a certified end of training. Expect this adapter to be replaced by a later step.",
|
| 15 |
+
"recorded": "2026-09-11"
|
| 16 |
+
}
|
gemma4_26b_a4b_190m/control/rlvr/direct/STEP.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 768,
|
| 3 |
+
"steps_published_at_source": [
|
| 4 |
+
16,
|
| 5 |
+
32,
|
| 6 |
+
64,
|
| 7 |
+
96,
|
| 8 |
+
128,
|
| 9 |
+
160,
|
| 10 |
+
192,
|
| 11 |
+
224,
|
| 12 |
+
256,
|
| 13 |
+
288,
|
| 14 |
+
320,
|
| 15 |
+
352,
|
| 16 |
+
384,
|
| 17 |
+
416,
|
| 18 |
+
448,
|
| 19 |
+
480,
|
| 20 |
+
512,
|
| 21 |
+
544,
|
| 22 |
+
576,
|
| 23 |
+
608,
|
| 24 |
+
640,
|
| 25 |
+
672,
|
| 26 |
+
704,
|
| 27 |
+
736,
|
| 28 |
+
768
|
| 29 |
+
],
|
| 30 |
+
"source": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1:rl-checkpoints/control-direct/step-768/",
|
| 31 |
+
"training_complete": true,
|
| 32 |
+
"rl_done": {
|
| 33 |
+
"status": "complete",
|
| 34 |
+
"max_steps": 768
|
| 35 |
+
},
|
| 36 |
+
"eval_endpoints_here": [
|
| 37 |
+
"control-agreement-step512",
|
| 38 |
+
"control-direct-holdoutclause-step768",
|
| 39 |
+
"control-direct-step768",
|
| 40 |
+
"control-pre_aft-step0"
|
| 41 |
+
],
|
| 42 |
+
"eval_coverage": "FULL campaign battery.",
|
| 43 |
+
"caveat": null,
|
| 44 |
+
"recorded": "2026-09-11"
|
| 45 |
+
}
|
gemma4_26b_a4b_190m/control/rlvr/thinking/STEP.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 256,
|
| 3 |
+
"steps_published_at_source": [
|
| 4 |
+
256
|
| 5 |
+
],
|
| 6 |
+
"source": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1:rl-checkpoints/control-thinking/step-256/",
|
| 7 |
+
"training_complete": false,
|
| 8 |
+
"rl_done": null,
|
| 9 |
+
"eval_endpoints_here": [
|
| 10 |
+
"control-thinking-holdoutclause-step256",
|
| 11 |
+
"control-thinking-step256"
|
| 12 |
+
],
|
| 13 |
+
"eval_coverage": "PARTIAL -- heldout-surface sweep at cap 12288 only. The rest of the battery was deliberately not run for the thinking legs; the absence of other endpoints is a decision, not a missing result.",
|
| 14 |
+
"caveat": "No RL_DONE receipt exists for this lineage: step 256 is the highest checkpoint PUBLISHED, not a certified end of training. Expect this adapter to be replaced by a later step.",
|
| 15 |
+
"recorded": "2026-09-11"
|
| 16 |
+
}
|