Upload folder using huggingface_hub
Browse files- final/adapter.pt +3 -0
- final/adapter_meta.json +147 -0
- final/recipe.yaml +54 -0
- final/training_state.pt +3 -0
final/adapter.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7d9930fc3f6995a21479593fe2135da898e1aeb3f9ba5c065bab202a5d104d59
|
| 3 |
+
size 3997609379
|
final/adapter_meta.json
ADDED
|
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"format_version": 1,
|
| 3 |
+
"base_model": "/content/models/Qwen3-0.6B",
|
| 4 |
+
"teacher_config": {
|
| 5 |
+
"name": "/content/models/Qwen3-4B-Instruct",
|
| 6 |
+
"num_hidden_layers": 36,
|
| 7 |
+
"num_attention_heads": 32,
|
| 8 |
+
"num_key_value_heads": 8,
|
| 9 |
+
"hidden_size": 2560,
|
| 10 |
+
"head_dim": 128
|
| 11 |
+
},
|
| 12 |
+
"projector_type": "complex",
|
| 13 |
+
"projector_config": {
|
| 14 |
+
"source_dim": 128,
|
| 15 |
+
"target_dim": 128,
|
| 16 |
+
"source_num_heads": 8,
|
| 17 |
+
"target_num_heads": 8,
|
| 18 |
+
"hidden_dim": 1024,
|
| 19 |
+
"intermediate_dim": 1024,
|
| 20 |
+
"num_layers": 3,
|
| 21 |
+
"pre_proj_layers": 3,
|
| 22 |
+
"dropout": 0.1,
|
| 23 |
+
"gate_init_temp": 1.0,
|
| 24 |
+
"gate_final_temp": 0.001,
|
| 25 |
+
"anneal_steps": 1929,
|
| 26 |
+
"zero_init": false
|
| 27 |
+
},
|
| 28 |
+
"layer_mapping": [
|
| 29 |
+
[
|
| 30 |
+
0,
|
| 31 |
+
8
|
| 32 |
+
],
|
| 33 |
+
[
|
| 34 |
+
1,
|
| 35 |
+
9
|
| 36 |
+
],
|
| 37 |
+
[
|
| 38 |
+
2,
|
| 39 |
+
10
|
| 40 |
+
],
|
| 41 |
+
[
|
| 42 |
+
3,
|
| 43 |
+
11
|
| 44 |
+
],
|
| 45 |
+
[
|
| 46 |
+
4,
|
| 47 |
+
12
|
| 48 |
+
],
|
| 49 |
+
[
|
| 50 |
+
5,
|
| 51 |
+
13
|
| 52 |
+
],
|
| 53 |
+
[
|
| 54 |
+
6,
|
| 55 |
+
14
|
| 56 |
+
],
|
| 57 |
+
[
|
| 58 |
+
7,
|
| 59 |
+
15
|
| 60 |
+
],
|
| 61 |
+
[
|
| 62 |
+
8,
|
| 63 |
+
16
|
| 64 |
+
],
|
| 65 |
+
[
|
| 66 |
+
9,
|
| 67 |
+
17
|
| 68 |
+
],
|
| 69 |
+
[
|
| 70 |
+
10,
|
| 71 |
+
18
|
| 72 |
+
],
|
| 73 |
+
[
|
| 74 |
+
11,
|
| 75 |
+
19
|
| 76 |
+
],
|
| 77 |
+
[
|
| 78 |
+
12,
|
| 79 |
+
20
|
| 80 |
+
],
|
| 81 |
+
[
|
| 82 |
+
13,
|
| 83 |
+
21
|
| 84 |
+
],
|
| 85 |
+
[
|
| 86 |
+
14,
|
| 87 |
+
22
|
| 88 |
+
],
|
| 89 |
+
[
|
| 90 |
+
15,
|
| 91 |
+
23
|
| 92 |
+
],
|
| 93 |
+
[
|
| 94 |
+
16,
|
| 95 |
+
24
|
| 96 |
+
],
|
| 97 |
+
[
|
| 98 |
+
17,
|
| 99 |
+
25
|
| 100 |
+
],
|
| 101 |
+
[
|
| 102 |
+
18,
|
| 103 |
+
26
|
| 104 |
+
],
|
| 105 |
+
[
|
| 106 |
+
19,
|
| 107 |
+
27
|
| 108 |
+
],
|
| 109 |
+
[
|
| 110 |
+
20,
|
| 111 |
+
28
|
| 112 |
+
],
|
| 113 |
+
[
|
| 114 |
+
21,
|
| 115 |
+
29
|
| 116 |
+
],
|
| 117 |
+
[
|
| 118 |
+
22,
|
| 119 |
+
30
|
| 120 |
+
],
|
| 121 |
+
[
|
| 122 |
+
23,
|
| 123 |
+
31
|
| 124 |
+
],
|
| 125 |
+
[
|
| 126 |
+
24,
|
| 127 |
+
32
|
| 128 |
+
],
|
| 129 |
+
[
|
| 130 |
+
25,
|
| 131 |
+
33
|
| 132 |
+
],
|
| 133 |
+
[
|
| 134 |
+
26,
|
| 135 |
+
34
|
| 136 |
+
],
|
| 137 |
+
[
|
| 138 |
+
27,
|
| 139 |
+
35
|
| 140 |
+
]
|
| 141 |
+
],
|
| 142 |
+
"step": 300,
|
| 143 |
+
"created": "2026-04-15T22:40:37.219231+00:00",
|
| 144 |
+
"num_projectors": 28,
|
| 145 |
+
"params_per_projector": 35688466,
|
| 146 |
+
"total_params": 999277048
|
| 147 |
+
}
|
final/recipe.yaml
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: Qwen/Qwen3-0.6B
|
| 2 |
+
teacher_model: Qwen/Qwen3-4B-Instruct-2507
|
| 3 |
+
layer_mapping: last_aligned
|
| 4 |
+
|
| 5 |
+
projector:
|
| 6 |
+
type: complex
|
| 7 |
+
hidden_dim: 1024
|
| 8 |
+
intermediate_dim: 1024
|
| 9 |
+
num_layers: 3
|
| 10 |
+
pre_proj_layers: 3
|
| 11 |
+
dropout: 0.1
|
| 12 |
+
gate_init_temp: 1.0
|
| 13 |
+
gate_final_temp: 0.001
|
| 14 |
+
anneal_steps: 1929
|
| 15 |
+
zero_init: false
|
| 16 |
+
|
| 17 |
+
training:
|
| 18 |
+
batch_size: 5
|
| 19 |
+
grad_accum: 52
|
| 20 |
+
max_steps: 300
|
| 21 |
+
lr: 1.0e-4
|
| 22 |
+
weight_decay: 0.01
|
| 23 |
+
warmup_ratio: 0.1
|
| 24 |
+
max_grad_norm: 1.0
|
| 25 |
+
max_length: 2048
|
| 26 |
+
scheduler: linear
|
| 27 |
+
compile_base: false
|
| 28 |
+
seed: 42
|
| 29 |
+
|
| 30 |
+
data:
|
| 31 |
+
dataset: teknium/OpenHermes-2.5
|
| 32 |
+
max_samples: 500000
|
| 33 |
+
train_ratio: 0.99
|
| 34 |
+
packing: false
|
| 35 |
+
num_workers: 4
|
| 36 |
+
pin_memory: true
|
| 37 |
+
|
| 38 |
+
checkpointing:
|
| 39 |
+
output_dir: checkpoints/qwen3_4b_complex
|
| 40 |
+
save_every: 50
|
| 41 |
+
verify_saves: true
|
| 42 |
+
hf_repo: MR0ssi/gunk-complex-4b
|
| 43 |
+
|
| 44 |
+
eval:
|
| 45 |
+
benchmarks:
|
| 46 |
+
- mmlu-redux
|
| 47 |
+
- ai2-arc
|
| 48 |
+
- openbookqa
|
| 49 |
+
- mmlu-pro
|
| 50 |
+
eval_every: 50
|
| 51 |
+
batch_size: 32
|
| 52 |
+
|
| 53 |
+
logging:
|
| 54 |
+
log_every: 10
|
final/training_state.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:288efa0924aeb88b2a469f72fdc856ad7cdd13ad8c3d145519fa879ee9eb8072
|
| 3 |
+
size 7995655565
|