Justin Brown commited on
Commit
9880b25
·
verified ·
1 Parent(s): c2f9ca0

Add checkpoint (fineweb): ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt

Browse files
README.md CHANGED
@@ -5,28 +5,30 @@ tags:
5
  - asa
6
  - asm
7
  - language-model
8
- - wikitext
9
  library_name: pytorch
10
  ---
11
 
12
- # ASA-ASM (Wikitext-103 raw) — Checkpoint
13
 
14
- This repo contains a checkpoint for an **Addressed State Model (ASM)** built from **Addressed State Attention (ASA)**.
15
 
16
- ## What's included
17
- - `/ASA_ASM_wt103-rawv1_gpt2_T1024_L21_D384_H8_K16_M32_ropek1_alibi1_gamma1_step75000_best.pt`: training checkpoint (PyTorch)
18
- - `config.json` (if present): training config serialized from checkpoint
19
- - `metadata.json`: SHA256 + provenance info (GitHub source pointer)
 
 
 
 
 
 
20
 
21
  ## Provenance
22
- - Code: https://github.com/digitaldaimyo/ASA
23
- - Revision: `codex/implement-runnable-and-releasable-setup`
24
 
25
- ## Quick load (raw checkpoint)
26
- ```python
27
- import torch
28
- ckpt = torch.load("ASA_ASM_wt103-rawv1_gpt2_T1024_L21_D384_H8_K16_M32_ropek1_alibi1_gamma1_step75000_best.pt", map_location="cpu")
29
- print(ckpt.keys())
30
 
31
  ## Notes
32
- This is a raw training artifact. A lightweight inference wrapper / conversion can be added later.
 
 
 
5
  - asa
6
  - asm
7
  - language-model
 
8
  library_name: pytorch
9
  ---
10
 
11
+ # ASA / ASM Checkpoints (Bucket Repo)
12
 
13
+ This repository stores **training checkpoints** for Addressed State Models (ASM) built from Addressed State Attention (ASA).
14
 
15
+ ## Repository layout
16
+
17
+ - `checkpoints/<dataset>/...pt`
18
+ Raw PyTorch checkpoints (training artifacts).
19
+
20
+ - `configs/<dataset>/<checkpoint>.config.json`
21
+ Training config extracted from the checkpoint (when present).
22
+
23
+ - `metadata/<dataset>/<checkpoint>.metadata.json`
24
+ SHA256 + file size + provenance pointers.
25
 
26
  ## Provenance
 
 
27
 
28
+ - Code: https://github.com/digitaldaimyo/ASA
29
+ - Revision: `main`
 
 
 
30
 
31
  ## Notes
32
+
33
+ - These are raw training artifacts, primarily intended for research and reproduction.
34
+ - New uploads may overwrite **rolling pointers** (e.g. `last.pt`, `best.pt`), but the Hub preserves history by commit.
checkpoints/fineweb/ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13e44f9f44d6ed669099918f1efa0440849b83ab374293e027887a35c385b055
3
+ size 2889860681
configs/fineweb/ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt.config.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "HuggingFaceFW/fineweb",
3
+ "dataset_config": "sample-10BT",
4
+ "tokenizer_name": "gpt2",
5
+ "max_seq_len": 1024,
6
+ "stride_frac_val": 0.5,
7
+ "seed": 1337,
8
+ "micro_batch_size": 2,
9
+ "grad_accum_steps": 16,
10
+ "train_samples_target": 800000,
11
+ "val_samples_target": 25000,
12
+ "batch_size": 4,
13
+ "learning_rate": 0.0003,
14
+ "weight_decay": 0.01,
15
+ "betas": [
16
+ 0.9,
17
+ 0.95
18
+ ],
19
+ "grad_clip": 1.0,
20
+ "warmup_steps": 300,
21
+ "total_steps": 50000,
22
+ "eval_interval": 1000,
23
+ "log_interval": 100,
24
+ "vocab_size": 50257,
25
+ "embed_dim": 1024,
26
+ "num_layers": 15,
27
+ "num_heads": 16,
28
+ "num_slots": 32,
29
+ "mlp_ratio": 4.0,
30
+ "dropout": 0.1,
31
+ "tie_weights": true,
32
+ "read_temperature": 1.0,
33
+ "write_temperature": 1.0,
34
+ "slot_dropout": 0.05,
35
+ "state_fp32": true,
36
+ "normalize_k": false,
37
+ "use_abs_pos": false,
38
+ "use_rope_keys": true,
39
+ "rope_base": 10000.0,
40
+ "use_alibi_write": true,
41
+ "alibi_strength_init": 0.1,
42
+ "learn_alibi_strength": true,
43
+ "min_strength": 0.0,
44
+ "use_content_read": true,
45
+ "content_read_init": -4.0,
46
+ "content_read_max_gamma": 3.0,
47
+ "use_slotspace_refine": true,
48
+ "slotspace_dim": 16,
49
+ "slotspace_gate_init": -4.0,
50
+ "slotspace_dropout": 0.05,
51
+ "slotspace_signed_weights": true,
52
+ "use_rope_slotspace": true,
53
+ "rope_base_slotspace": 100000.0,
54
+ "write_chunk_size": 1024,
55
+ "enable_compiled": true,
56
+ "eval_max_batches": 150,
57
+ "analytics_last_k": 4,
58
+ "output_dir": "./drive/MyDrive/asm_outputs",
59
+ "tag": "asm_fineweb_1024t_1024d_16h_16sd_32s_1024cs_15l",
60
+ "cache_dir": "./drive/MyDrive/asm_caches/fineweb",
61
+ "val_windows_cache": "./drive/MyDrive/asm_nlp/val_cache_fineweb_u32_windows_1024.u32tok"
62
+ }
metadata/fineweb/ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt.metadata.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "repo_id": "DigitalShogun/ASA-ASM-wikitext103-raw",
3
+ "dataset": "fineweb",
4
+ "exported_at_utc": "2026-01-30T16:21:48.689906+00:00",
5
+ "checkpoint_filename": "ASA_ASM_fineweb_T1024_L15_D1024_H16_K16_S32_step17500_last.pt",
6
+ "checkpoint_sha256": "13e44f9f44d6ed669099918f1efa0440849b83ab374293e027887a35c385b055",
7
+ "checkpoint_bytes": 2889860681,
8
+ "github_repo": "https://github.com/digitaldaimyo/ASA",
9
+ "github_revision": "main",
10
+ "train_config_in_ckpt": {
11
+ "dataset_name": "HuggingFaceFW/fineweb",
12
+ "dataset_config": "sample-10BT",
13
+ "tokenizer_name": "gpt2",
14
+ "max_seq_len": 1024,
15
+ "stride_frac_val": 0.5,
16
+ "seed": 1337,
17
+ "micro_batch_size": 2,
18
+ "grad_accum_steps": 16,
19
+ "train_samples_target": 800000,
20
+ "val_samples_target": 25000,
21
+ "batch_size": 4,
22
+ "learning_rate": 0.0003,
23
+ "weight_decay": 0.01,
24
+ "betas": [
25
+ 0.9,
26
+ 0.95
27
+ ],
28
+ "grad_clip": 1.0,
29
+ "warmup_steps": 300,
30
+ "total_steps": 50000,
31
+ "eval_interval": 1000,
32
+ "log_interval": 100,
33
+ "vocab_size": 50257,
34
+ "embed_dim": 1024,
35
+ "num_layers": 15,
36
+ "num_heads": 16,
37
+ "num_slots": 32,
38
+ "mlp_ratio": 4.0,
39
+ "dropout": 0.1,
40
+ "tie_weights": true,
41
+ "read_temperature": 1.0,
42
+ "write_temperature": 1.0,
43
+ "slot_dropout": 0.05,
44
+ "state_fp32": true,
45
+ "normalize_k": false,
46
+ "use_abs_pos": false,
47
+ "use_rope_keys": true,
48
+ "rope_base": 10000.0,
49
+ "use_alibi_write": true,
50
+ "alibi_strength_init": 0.1,
51
+ "learn_alibi_strength": true,
52
+ "min_strength": 0.0,
53
+ "use_content_read": true,
54
+ "content_read_init": -4.0,
55
+ "content_read_max_gamma": 3.0,
56
+ "use_slotspace_refine": true,
57
+ "slotspace_dim": 16,
58
+ "slotspace_gate_init": -4.0,
59
+ "slotspace_dropout": 0.05,
60
+ "slotspace_signed_weights": true,
61
+ "use_rope_slotspace": true,
62
+ "rope_base_slotspace": 100000.0,
63
+ "write_chunk_size": 1024,
64
+ "enable_compiled": true,
65
+ "eval_max_batches": 150,
66
+ "analytics_last_k": 4,
67
+ "output_dir": "./drive/MyDrive/asm_outputs",
68
+ "tag": "asm_fineweb_1024t_1024d_16h_16sd_32s_1024cs_15l",
69
+ "cache_dir": "./drive/MyDrive/asm_caches/fineweb",
70
+ "val_windows_cache": "./drive/MyDrive/asm_nlp/val_cache_fineweb_u32_windows_1024.u32tok"
71
+ },
72
+ "ckpt_step_inferred": 17500,
73
+ "ckpt_step_fields": {
74
+ "global_step": null,
75
+ "step": 17500,
76
+ "train_step": null,
77
+ "epoch": null,
78
+ "iter": null,
79
+ "iteration": null
80
+ },
81
+ "ckpt_keys": [
82
+ "cfg",
83
+ "model",
84
+ "opt",
85
+ "step",
86
+ "best_val"
87
+ ]
88
+ }