davidheineman commited on
Commit
a84df80
·
verified ·
1 Parent(s): 566d7ea

Archive completed pilot-2m__8m__arithmetic_simple

Browse files
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/ancestry.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "rl_initializer": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/pilot-2m/runs/8m/arithmetic_simple/sft.pt",
3
+ "terminal_sft_weights": "3f6a19449bb6344fe4a1a870592ee780965a3060c38116ebdc1e2ad3877c0be8",
4
+ "full_sft_budget": 2097152,
5
+ "intermediate_checkpoint_selection": false,
6
+ "hint_distribution": "Independent uniform integer 1..16 per query",
7
+ "adaptive_calibration": false,
8
+ "samples_per_query": 32
9
+ }
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/archive-manifest.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "protocol": "57ad2f146574a01a39a5a991770f40c0d22eae416970b0064a18bf32e574daaa",
3
+ "case": {
4
+ "key": "pilot-2m__8m__arithmetic_simple",
5
+ "folder": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/pilot-2m/runs/8m/arithmetic_simple",
6
+ "reuse": false
7
+ },
8
+ "files": {
9
+ "sft.pt": {
10
+ "size": 32060780,
11
+ "sha256": "89c95f504e051811d04a23fea6e158f68fd201759c56c60e0bd647416b624acb"
12
+ },
13
+ "rl.pt": {
14
+ "size": 32060669,
15
+ "sha256": "7239a36552f6c61806bd1b6150b201d201c8cbb1096d9c6fb6699cbf42ff6312"
16
+ }
17
+ },
18
+ "complete": {
19
+ "task": "arithmetic_simple",
20
+ "model": "8m",
21
+ "sft_tokens": 2097152,
22
+ "sft_budget": 2097152,
23
+ "rl_rollouts": 65536,
24
+ "rl_budget": 65536,
25
+ "terminal_sft_hash": "3f6a19449bb6344fe4a1a870592ee780965a3060c38116ebdc1e2ad3877c0be8",
26
+ "final_rl_hash": "d7c4813eba4a5140b117dff1d0e4f6dd6bafd271190bb214da5e98f9592f20a1",
27
+ "sft_seconds": 75.65454360470176,
28
+ "rl_seconds": 389.4870395511389,
29
+ "no_checkpoint_selection": true,
30
+ "evaluation_generations": 10752,
31
+ "calibration_generations": 0,
32
+ "hint_distribution": "uniform 1..16 per query",
33
+ "samples_per_query": 32,
34
+ "development_episodes": 64,
35
+ "development_unique_groups": 64
36
+ }
37
+ }
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/complete.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "task": "arithmetic_simple",
3
+ "model": "8m",
4
+ "sft_tokens": 2097152,
5
+ "sft_budget": 2097152,
6
+ "rl_rollouts": 65536,
7
+ "rl_budget": 65536,
8
+ "terminal_sft_hash": "3f6a19449bb6344fe4a1a870592ee780965a3060c38116ebdc1e2ad3877c0be8",
9
+ "final_rl_hash": "d7c4813eba4a5140b117dff1d0e4f6dd6bafd271190bb214da5e98f9592f20a1",
10
+ "sft_seconds": 75.65454360470176,
11
+ "rl_seconds": 389.4870395511389,
12
+ "no_checkpoint_selection": true,
13
+ "evaluation_generations": 10752,
14
+ "calibration_generations": 0,
15
+ "hint_distribution": "uniform 1..16 per query",
16
+ "samples_per_query": 32,
17
+ "development_episodes": 64,
18
+ "development_unique_groups": 64
19
+ }
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/request.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "task": "arithmetic_simple",
4
+ "model": "8m",
5
+ "root": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/pilot-2m",
6
+ "checkpoint": "/nlp/scr/dhei/nanogpt-ladder-v3/runs/base/8m-240353280/seed-1337/model.pt",
7
+ "output": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/pilot-2m/runs/8m/arithmetic_simple",
8
+ "sft_tokens": 2097152,
9
+ "rl_rollouts": 65536,
10
+ "samples": 32,
11
+ "groups": 4,
12
+ "batch_tokens": 4096,
13
+ "micro_tokens": 2048,
14
+ "rl_micro_tokens": 8192,
15
+ "rl_micro_batch": 16,
16
+ "sft_lr": 0.0003,
17
+ "sft_hidden_lr": 0.0005,
18
+ "rl_lr": 3e-05,
19
+ "rl_hidden_lr": 5e-05,
20
+ "conditioning": "hinted-v2",
21
+ "qk_rms": 0.03,
22
+ "qk_update_scale": 0.1,
23
+ "kl": 0.01,
24
+ "clip": 0.2,
25
+ "eval_queries": 64,
26
+ "eval_samples": 8,
27
+ "eval_batch": 2,
28
+ "interval": 4096,
29
+ "recovery_interval": 4096,
30
+ "recovery_seconds": 120.0,
31
+ "workers": 8,
32
+ "judged_backend": "skywork",
33
+ "judge_cache": "",
34
+ "threads": 2,
35
+ "seed": 2718,
36
+ "device": "auto",
37
+ "wandb_project": "tiny-rl-refined-math-v1",
38
+ "wandb_entity": "",
39
+ "wandb_mode": "online",
40
+ "smoke": false
41
+ },
42
+ "checkpoint_sha256": "d61387eac987925fb648d9d670d2e13994c574e2913f81880ed967da51103cb5",
43
+ "sft_sha256": "572f39ce353ccfea586bb37044cfad0497f28d0baf93e213f008c8f71ac851e9",
44
+ "trainer_sha256": "c5e0b0ad8cafb1df841fbe0ee622d06f605c175b462d37a69222498d284a79ec",
45
+ "policy_sha256": "4038b4f4897dcbf253625ee3e0646540c3f8b7c5a734e9f1957132e5660cdc83",
46
+ "metrics_sha256": "75ae4dc056330c597f18aa019d28bfa315dc84e99f7e89df5a2fe47b25699b2c",
47
+ "data_sha256": "0c95b6331b6c57f897a16ce5b2c4328428923661c63092919d84ab6fb631f012",
48
+ "split_sha256": {
49
+ "rl": "4ad67fa0c1a2f316ea9ca73aea1146532370eada3bd6fe088d510b0231241d96",
50
+ "dev": "08b9165b328f95e990b57e225e4972fb5f7528d24cf4752a8d789b8319fce6a3",
51
+ "test": "e38f6dadd11b931f52578a94a5089db7faaff8366524d6bf93e0e861029574c3"
52
+ },
53
+ "environment_sha256": "13356da2dea775fa0df14959b97e5bd079b25929d35f933a82157c6934d5ed21"
54
+ }
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/rl.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7239a36552f6c61806bd1b6150b201d201c8cbb1096d9c6fb6699cbf42ff6312
3
+ size 32060669
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/sft.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89c95f504e051811d04a23fea6e158f68fd201759c56c60e0bd647416b624acb
3
+ size 32060780
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/startup.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "raw_weights": "66c870dcf87e9d97b6f3040a50f5f120a0c3e87e3d5f265cfbf4d323f293832b",
3
+ "conditioning": {
4
+ "recipe": "Latest functional-hints-v2 shared Q/K and residual conditioning, applied before task SFT; this changes the raw pretrained function. No warmup or additional tokens.",
5
+ "residual_projection_scale": 0.005598449917504264,
6
+ "qk_rms": 0.03,
7
+ "initialized_weights": "852c267929b1a5e40a1a0fa9286170dad27e8956c9a3e22dcb1fd5e673d3cc9b"
8
+ },
9
+ "parameters": 8011776,
10
+ "hardware": "NVIDIA A100-SXM4-80GB",
11
+ "torch_version": "2.10.0+cu128",
12
+ "cuda_runtime": "12.8",
13
+ "torch_path": "/juice5b/u/dhei/tatsu/nano-omni/.venv/lib/python3.12/site-packages/torch/__init__.py",
14
+ "job": "18429746",
15
+ "restart": "0"
16
+ }
cases/57ad2f146574a01a/pilot-2m__8m__arithmetic_simple/terminal-sft.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "tokens": 2097152,
3
+ "budget": 2097152,
4
+ "weights_sha256": "3f6a19449bb6344fe4a1a870592ee780965a3060c38116ebdc1e2ad3877c0be8",
5
+ "seconds": 75.65454360470176,
6
+ "role": "Only terminal full-budget SFT weights initialize RL"
7
+ }