Archive completed original-v4-full-math__64m__arithmetic

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/ancestry.json

@@ -0,0 +1,9 @@
+{
+  "rl_initializer": "/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4/runs/64m/arithmetic/sft.pt",
+  "terminal_sft_weights": "336313de99a040527964ca7a8e56e4da219a833001ba6160ebbe05549e00ac93",
+  "full_sft_budget": 16777216,
+  "intermediate_checkpoint_selection": false,
+  "hint_distribution": "Independent uniform integer 1..16 per query",
+  "adaptive_calibration": false,
+  "samples_per_query": 32
+}

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/archive-manifest.json

@@ -0,0 +1,37 @@
+{
+  "protocol": "33898773f348926fcdfad0e474ce574985e92054105d1f76c4392d32ac659de9",
+  "case": {
+    "key": "original-v4-full-math__64m__arithmetic",
+    "folder": "/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4/runs/64m/arithmetic",
+    "reuse": false
+  },
+  "files": {
+    "sft.pt": {
+      "size": 254038368,
+      "sha256": "d76f6bd7d495452b5830b2a9f7640a48c634e70d1d913cc1a6a5df4b6482116f"
+    },
+    "rl.pt": {
+      "size": 254038237,
+      "sha256": "93952e5b2398a0e1c70731228a8b29e21c1cb27f35dd8d1388abf2396e0201a9"
+    }
+  },
+  "complete": {
+    "task": "arithmetic",
+    "model": "64m",
+    "sft_tokens": 16777216,
+    "sft_budget": 16777216,
+    "rl_rollouts": 65536,
+    "rl_budget": 65536,
+    "terminal_sft_hash": "336313de99a040527964ca7a8e56e4da219a833001ba6160ebbe05549e00ac93",
+    "final_rl_hash": "89abb499e3710e8594a464546f4f74ec54600f99a2bdc7ca884aa0c4f0fe76fb",
+    "sft_seconds": 2408.8948959279805,
+    "rl_seconds": 1584.5415151566267,
+    "no_checkpoint_selection": true,
+    "evaluation_generations": 12288,
+    "calibration_generations": 0,
+    "hint_distribution": "uniform 1..16 per query",
+    "samples_per_query": 32,
+    "development_episodes": 64,
+    "development_unique_groups": 64
+  }
+}

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/complete.json

@@ -0,0 +1,19 @@
+{
+  "task": "arithmetic",
+  "model": "64m",
+  "sft_tokens": 16777216,
+  "sft_budget": 16777216,
+  "rl_rollouts": 65536,
+  "rl_budget": 65536,
+  "terminal_sft_hash": "336313de99a040527964ca7a8e56e4da219a833001ba6160ebbe05549e00ac93",
+  "final_rl_hash": "89abb499e3710e8594a464546f4f74ec54600f99a2bdc7ca884aa0c4f0fe76fb",
+  "sft_seconds": 2408.8948959279805,
+  "rl_seconds": 1584.5415151566267,
+  "no_checkpoint_selection": true,
+  "evaluation_generations": 12288,
+  "calibration_generations": 0,
+  "hint_distribution": "uniform 1..16 per query",
+  "samples_per_query": 32,
+  "development_episodes": 64,
+  "development_unique_groups": 64
+}

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/request.json

@@ -0,0 +1,52 @@
+{
+  "config": {
+    "task": "arithmetic",
+    "model": "64m",
+    "root": "/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4",
+    "checkpoint": "/nlp/scr/dhei/nanogpt-ladder-v3/runs/base/64m-1905131520/seed-1337/model.pt",
+    "output": "/nlp/scr/dhei/tiny-rl/experiments/refined-tasks-v4/runs/64m/arithmetic",
+    "sft_tokens": 16777216,
+    "rl_rollouts": 65536,
+    "samples": 32,
+    "groups": 4,
+    "batch_tokens": 4096,
+    "micro_tokens": 2048,
+    "rl_micro_tokens": 8192,
+    "rl_micro_batch": 16,
+    "sft_lr": 0.0003,
+    "sft_hidden_lr": 0.0005,
+    "rl_lr": 3e-05,
+    "rl_hidden_lr": 5e-05,
+    "conditioning": "hinted-v2",
+    "qk_rms": 0.03,
+    "qk_update_scale": 0.1,
+    "kl": 0.01,
+    "clip": 0.2,
+    "eval_queries": 64,
+    "eval_samples": 8,
+    "eval_batch": 2,
+    "interval": 4096,
+    "recovery_interval": 4096,
+    "recovery_seconds": 120.0,
+    "workers": 8,
+    "threads": 2,
+    "seed": 2718,
+    "device": "auto",
+    "wandb_project": "tiny-rl-refined-tasks-v4",
+    "wandb_entity": "",
+    "wandb_mode": "online",
+    "smoke": false
+  },
+  "checkpoint_sha256": "ca3ac0c59db71de1bd56c29d6f3b51a419085084e0ab5d421da7f9debf47e1b9",
+  "sft_sha256": "6d0136f06866377925d637d3ec3f6dfd1ff09e10d77c62fec2711088e1b3cd01",
+  "trainer_sha256": "b58f90881881859eab07fba169ea7d10729f03b4b0b6d076d683fdc8af8658aa",
+  "policy_sha256": "4038b4f4897dcbf253625ee3e0646540c3f8b7c5a734e9f1957132e5660cdc83",
+  "metrics_sha256": "d96631d6def4c80ab05a44d18c8ef2ff0b0574c2bae1ff1016f64759b60a3cda",
+  "data_sha256": "f3d27ec4caf49490b7be3de0702bd3c5298a6b55ed73361c83e307a820042ed3",
+  "split_sha256": {
+    "rl": "450e23c846b4827959fd1d1a8f5e9994038543ff56d56574159c9276b8cfba90",
+    "dev": "115c4f5f71241592486567af9ae0066c05e37c54bfb8a3d856b716751ecf1ddd",
+    "test": "37ff561ae25cbbcabdcc76be6870e868d37d3cca48665ff593ead33d5899c056"
+  },
+  "environment_sha256": "13356da2dea775fa0df14959b97e5bd079b25929d35f933a82157c6934d5ed21"
+}

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/rl.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93952e5b2398a0e1c70731228a8b29e21c1cb27f35dd8d1388abf2396e0201a9
+size 254038237

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/sft.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d76f6bd7d495452b5830b2a9f7640a48c634e70d1d913cc1a6a5df4b6482116f
+size 254038368

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/startup.json

@@ -0,0 +1,16 @@
+{
+  "raw_weights": "1d909208ee8b57a04c2382e2b0c9c17c21ff2f0a841613a42f5c03ff87f01cc1",
+  "conditioning": {
+    "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.",
+    "residual_projection_scale": 0.002387429619854719,
+    "qk_rms": 0.03,
+    "initialized_weights": "4b3440c3df737911f597e38ea104427ea8e020b3255fad26e31287a5d5daaf02"
+  },
+  "parameters": 63504384,
+  "hardware": "NVIDIA GeForce GTX 1080 Ti",
+  "torch_version": "2.10.0+cu126",
+  "cuda_runtime": "12.6",
+  "torch_path": "/juice5b/scr5b/dhei/tiny-rl/experiments/refined-tasks-v4/runtime-cu126/torch/__init__.py",
+  "job": "18425702",
+  "restart": "0"
+}

cases/33898773f348926f/original-v4-full-math__64m__arithmetic/terminal-sft.json

@@ -0,0 +1,7 @@
+{
+  "tokens": 16777216,
+  "budget": 16777216,
+  "weights_sha256": "336313de99a040527964ca7a8e56e4da219a833001ba6160ebbe05549e00ac93",
+  "seconds": 2408.8948959279805,
+  "role": "Only terminal full-budget SFT weights initialize RL"
+}