step 718 shard 44
Browse files- README.md +2 -7
- config.json +53 -8
- model.safetensors +1 -1
README.md
CHANGED
|
@@ -1,10 +1,5 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
library_name: pytorch
|
| 4 |
-
---
|
| 5 |
-
|
| 6 |
# lobotomy-1b
|
| 7 |
|
| 8 |
-
21-layer ~1B student
|
| 9 |
|
| 10 |
-
step=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# lobotomy-1b
|
| 2 |
|
| 3 |
+
21-layer ~1B student (Qwen3.8-27B pooled 64→21, downproj 5120→1536).
|
| 4 |
|
| 5 |
+
step=718 pass=
|
config.json
CHANGED
|
@@ -1,6 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"
|
| 3 |
-
"cfg": {
|
| 4 |
"vocab_size": 248320,
|
| 5 |
"hidden_size": 1536,
|
| 6 |
"n_layers": 21,
|
|
@@ -19,10 +18,56 @@
|
|
| 19 |
"logit_softcap": 15.0,
|
| 20 |
"mtp": 1
|
| 21 |
},
|
| 22 |
-
"
|
| 23 |
-
"
|
| 24 |
-
"
|
| 25 |
-
"
|
| 26 |
-
"
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"model": {
|
|
|
|
| 3 |
"vocab_size": 248320,
|
| 4 |
"hidden_size": 1536,
|
| 5 |
"n_layers": 21,
|
|
|
|
| 18 |
"logit_softcap": 15.0,
|
| 19 |
"mtp": 1
|
| 20 |
},
|
| 21 |
+
"train": {
|
| 22 |
+
"seq_len": 256,
|
| 23 |
+
"batch_size": 1,
|
| 24 |
+
"max_steps": 8000,
|
| 25 |
+
"warmup_steps": 200,
|
| 26 |
+
"cooldown_frac": 0.45,
|
| 27 |
+
"decay_steps": 40,
|
| 28 |
+
"lr": 0.0003,
|
| 29 |
+
"muon_lr": 0.02,
|
| 30 |
+
"embed_lr": 0.0006,
|
| 31 |
+
"head_lr": 0.0003,
|
| 32 |
+
"muon_momentum_warmup": 200,
|
| 33 |
+
"weight_decay": 0.1,
|
| 34 |
+
"seq_schedule": [],
|
| 35 |
+
"grad_clip": 1.0,
|
| 36 |
+
"optimizer": "muon",
|
| 37 |
+
"bf16": true,
|
| 38 |
+
"compile": false,
|
| 39 |
+
"seed": 0,
|
| 40 |
+
"log_every": 10,
|
| 41 |
+
"save_every": 200,
|
| 42 |
+
"out_dir": "checkpoints/1b"
|
| 43 |
+
},
|
| 44 |
+
"distill": {
|
| 45 |
+
"enabled": true,
|
| 46 |
+
"teacher": "Qwen/Qwen3.8-27B",
|
| 47 |
+
"mode": "hub",
|
| 48 |
+
"cache_dir": "data/teacher",
|
| 49 |
+
"load_in_4bit": true,
|
| 50 |
+
"weight": 0.0,
|
| 51 |
+
"ce_weight": 0.1,
|
| 52 |
+
"logit_weight": 1.0,
|
| 53 |
+
"loss": "cosine",
|
| 54 |
+
"layers": "grouped",
|
| 55 |
+
"pool": "last",
|
| 56 |
+
"pca_batches": 0,
|
| 57 |
+
"subsample": 0,
|
| 58 |
+
"max_seq": 256,
|
| 59 |
+
"temperature": 2.0,
|
| 60 |
+
"trust_remote_code": true,
|
| 61 |
+
"hub_repo": "avewright/qwen-code-soft-logits",
|
| 62 |
+
"hub_model": "avewright/lobotomy-1b",
|
| 63 |
+
"hub_private": true,
|
| 64 |
+
"shard_seqs": 16,
|
| 65 |
+
"max_local_gb": 6.0,
|
| 66 |
+
"local_dir": "data/teacher_shards",
|
| 67 |
+
"max_passes": 1,
|
| 68 |
+
"upload_every_shards": 5
|
| 69 |
+
},
|
| 70 |
+
"step": 718,
|
| 71 |
+
"shard_i": 44,
|
| 72 |
+
"pass": ""
|
| 73 |
}
|
model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 2698072840
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:197a32f24702283ba9d61395b24721d441ed6363561cbd1e56249dcd33f31ce2
|
| 3 |
size 2698072840
|