Publish verified 8loop-stage2-proposal-low-lr-300 recurrent weights
Browse files
8loop-stage2-proposal-low-lr-300/checkpoint.json
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"id": "8loop-stage2-proposal-low-lr-300",
|
| 3 |
+
"student_depth": 8,
|
| 4 |
+
"teacher_depth": 32,
|
| 5 |
+
"stage1_updates": 1000,
|
| 6 |
+
"stage2_updates": 300,
|
| 7 |
+
"checkpoint_step": 300,
|
| 8 |
+
"learning_rate": 2e-07,
|
| 9 |
+
"comparison_report": "reports/continuation_500_20261008/comparison.json",
|
| 10 |
+
"selection": "last fully evaluated matched checkpoint before user stop",
|
| 11 |
+
"evaluation": {
|
| 12 |
+
"questions": 128,
|
| 13 |
+
"final_matches": 48,
|
| 14 |
+
"strict_matches": 34,
|
| 15 |
+
"cap_hits": 2,
|
| 16 |
+
"split": "held-out GSM8K development",
|
| 17 |
+
"generation_cap": 256
|
| 18 |
+
},
|
| 19 |
+
"prefix_policy": "cached teacher32-generated answer prefixes",
|
| 20 |
+
"format": "loop-distil-recurrent-bf16-v1",
|
| 21 |
+
"base_model": "tomg-group-umd/huginn-0125",
|
| 22 |
+
"base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
|
| 23 |
+
"weights_sha256": "67197f0bf7b28d2b071b42e6cb56f2a7b71768845173745144187246ea4dc787",
|
| 24 |
+
"source_master_sha256": "a7598488d52fb8818212b35939bfd9ccff4ed847d9879f3322ab5750f151fe17",
|
| 25 |
+
"tensor_count": 37,
|
| 26 |
+
"parameters": 1637349120,
|
| 27 |
+
"weights_bytes": 3274702592,
|
| 28 |
+
"dtype": "bfloat16",
|
| 29 |
+
"exact_bf16_compute_roundtrip": true,
|
| 30 |
+
"optimizer_included": false,
|
| 31 |
+
"source_checkpoint_filename": "step_0300.pt",
|
| 32 |
+
"note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints."
|
| 33 |
+
}
|
8loop-stage2-proposal-low-lr-300/recurrent.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:67197f0bf7b28d2b071b42e6cb56f2a7b71768845173745144187246ea4dc787
|
| 3 |
+
size 3274702592
|