echoctx commited on
Commit
bdcfcd9
·
verified ·
1 Parent(s): ca5bc16

sn38 2016 unique nanochrono + quality SFT2

Browse files
README.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ tags:
4
+ - sn38
5
+ - chronollm
6
+ - causal-lm
7
+ ---
8
+
9
+ # sn38-chrono-2016-sft2
10
+
11
+ Round-7 / cutoff-2016 candidate for Bittensor SN38.
12
+
13
+ - Arch: `sn38-nanochrono`, 2,018,511,234 params
14
+ - Init: `anacoluthe89/chrono-2015@3062e6d0` then SV-jitter sigma=0.08 (SVD 0.019 vs init)
15
+ - Then completion-only SFT on 113 timeless incomplete-sentence pairs (no post-2016 dumps)
16
+ - Official SVD vs UID 131 still **0.01905 >= 0.01**
17
+ - Greedy generation_config; trust_remote_code=False
18
+ - Held-out completions are shorter and more factual than the 2015 champion on a 20-prompt set
19
+ - Proxy leak unknown-recognition rose (year-token items); TEE self-test not yet run
config.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "NanochronoForCausalLM"
4
+ ],
5
+ "attention_multiplier": 1.2,
6
+ "aux_gate_dim": 12,
7
+ "aux_layers": [
8
+ 1,
9
+ 3,
10
+ 5,
11
+ 7,
12
+ 9,
13
+ 11,
14
+ 13,
15
+ 15,
16
+ 17,
17
+ 19,
18
+ 21,
19
+ 23,
20
+ 25,
21
+ 27
22
+ ],
23
+ "bos_token_id": 32759,
24
+ "dtype": "bfloat16",
25
+ "eos_token_id": 32763,
26
+ "final_logit_softcapping": 15.0,
27
+ "hidden_size": 1792,
28
+ "intermediate_size": 7168,
29
+ "layer_types": [
30
+ "sliding_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "sliding_attention",
37
+ "full_attention",
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "full_attention",
42
+ "sliding_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "full_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "full_attention",
54
+ "sliding_attention",
55
+ "sliding_attention",
56
+ "sliding_attention",
57
+ "full_attention"
58
+ ],
59
+ "max_position_embeddings": 2048,
60
+ "mix_gate_dim": 24,
61
+ "model_type": "sn38-nanochrono",
62
+ "num_attention_heads": 14,
63
+ "num_hidden_layers": 28,
64
+ "num_key_value_heads": 14,
65
+ "pad_token_id": 32763,
66
+ "rope_theta": 100000.0,
67
+ "sliding_window": 512,
68
+ "tap_layer": 14,
69
+ "tie_word_embeddings": false,
70
+ "transformers_version": "5.15.0",
71
+ "use_cache": true,
72
+ "vocab_size": 32768
73
+ }
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 32759,
3
+ "do_sample": false,
4
+ "eos_token_id": 32763,
5
+ "max_new_tokens": 50,
6
+ "pad_token_id": 32763,
7
+ "transformers_version": "5.15.0"
8
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc04829ec1cd2d1e92211feafc151d35f9335f8a2a6a42d460bd304fbb88936f
3
+ size 4037045748
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<|bos|>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "eos_token": "<|assistant_end|>",
6
+ "is_local": true,
7
+ "local_files_only": false,
8
+ "model_max_length": 2048,
9
+ "pad_token": "<|assistant_end|>",
10
+ "tokenizer_class": "TokenizersBackend"
11
+ }
train_meta.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "init": "/data/ckpt/nano-sv-0.08",
3
+ "revision": null,
4
+ "steps": 120,
5
+ "lr": 2e-05,
6
+ "nparams": 2018511234,
7
+ "npairs": 113,
8
+ "note": "completion-only SFT on timeless quality pairs; no web dumps",
9
+ "last_loss": 0.4078323543071747,
10
+ "step": 120,
11
+ "saved_at": 1786735651.4264104
12
+ }