Archive completed projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/ancestry.json

@@ -0,0 +1,9 @@
+{
+  "rl_initializer": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/projection-only-v1/runs/pretrained-attention-scaled-projections/32m/partial_gsm8k/sft.pt",
+  "terminal_sft_weights": "a37df7f3574a399036bfd9c9771a1ea1e1e5e56365bd48c9b33392c95dc75964",
+  "full_sft_budget": 2097152,
+  "intermediate_checkpoint_selection": false,
+  "hint_distribution": "Independent uniform integer 1..16 per query",
+  "adaptive_calibration": false,
+  "samples_per_query": 32
+}

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/archive-manifest.json

@@ -0,0 +1,37 @@
+{
+  "protocol": "202f6bc88a76dd3d0fe65c94035d02f00a9f8c308a32864f7138a5c8ae8d5709",
+  "case": {
+    "key": "projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections",
+    "folder": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/projection-only-v1/runs/pretrained-attention-scaled-projections/32m/partial_gsm8k",
+    "reuse": false
+  },
+  "files": {
+    "sft.pt": {
+      "size": 128309181,
+      "sha256": "a5969e051254e4a2ef78db286d91b9eecebf118b6ebf6118a3b20c810977bbbb"
+    },
+    "rl.pt": {
+      "size": 128309045,
+      "sha256": "f9c898d8bac3c9501e5bb047b0bdc7e570887c8f3c9b6b1db23c5d790051a66f"
+    }
+  },
+  "complete": {
+    "task": "partial_gsm8k",
+    "model": "32m",
+    "sft_tokens": 2097152,
+    "sft_budget": 2097152,
+    "rl_rollouts": 65536,
+    "rl_budget": 65536,
+    "terminal_sft_hash": "a37df7f3574a399036bfd9c9771a1ea1e1e5e56365bd48c9b33392c95dc75964",
+    "final_rl_hash": "b8559481dbed2a1f2f0eefc17ca3d841f08faa55e59b5aa8146d09cd955850cc",
+    "sft_seconds": 202.71570450719446,
+    "rl_seconds": 633.3891215119511,
+    "no_checkpoint_selection": true,
+    "evaluation_generations": 10752,
+    "calibration_generations": 0,
+    "hint_distribution": "uniform 1..16 per query",
+    "samples_per_query": 32,
+    "development_episodes": 64,
+    "development_unique_groups": 64
+  }
+}

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/complete.json

@@ -0,0 +1,19 @@
+{
+  "task": "partial_gsm8k",
+  "model": "32m",
+  "sft_tokens": 2097152,
+  "sft_budget": 2097152,
+  "rl_rollouts": 65536,
+  "rl_budget": 65536,
+  "terminal_sft_hash": "a37df7f3574a399036bfd9c9771a1ea1e1e5e56365bd48c9b33392c95dc75964",
+  "final_rl_hash": "b8559481dbed2a1f2f0eefc17ca3d841f08faa55e59b5aa8146d09cd955850cc",
+  "sft_seconds": 202.71570450719446,
+  "rl_seconds": 633.3891215119511,
+  "no_checkpoint_selection": true,
+  "evaluation_generations": 10752,
+  "calibration_generations": 0,
+  "hint_distribution": "uniform 1..16 per query",
+  "samples_per_query": 32,
+  "development_episodes": 64,
+  "development_unique_groups": 64
+}

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/request.json

@@ -0,0 +1,63 @@
+{
+  "config": {
+    "task": "partial_gsm8k",
+    "model": "32m",
+    "root": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/projection-only-v1",
+    "checkpoint": "/nlp/scr/dhei/nanogpt-ladder-v3/runs/base/32m-962134016/seed-1337/model.pt",
+    "output": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/projection-only-v1/runs/pretrained-attention-scaled-projections/32m/partial_gsm8k",
+    "sft_tokens": 2097152,
+    "rl_rollouts": 65536,
+    "samples": 32,
+    "groups": 4,
+    "batch_tokens": 512,
+    "micro_tokens": 2048,
+    "preserve_sft_context": false,
+    "rl_micro_tokens": 8192,
+    "rl_micro_batch": 16,
+    "sft_lr": 0.0003,
+    "sft_hidden_lr": 0.0005,
+    "rl_lr": 1e-05,
+    "rl_hidden_lr": 1.6666666666666667e-05,
+    "conditioning": "hinted-v2",
+    "optimizer": "relative",
+    "qk_rms": 0.03,
+    "qk_update_scale": 1.0,
+    "kl": 0.1,
+    "clip": 0.2,
+    "eval_queries": 64,
+    "eval_samples": 8,
+    "eval_batch": 2,
+    "sft_eval_interval": 0,
+    "interval": 4096,
+    "recovery_interval": 4096,
+    "recovery_seconds": 120.0,
+    "workers": 8,
+    "judged_backend": "skywork",
+    "judge_cache": "",
+    "chat_stop": "none",
+    "threads": 2,
+    "seed": 2718,
+    "device": "auto",
+    "wandb_project": "tiny-rl-refined-math-v1",
+    "wandb_entity": "david-heineman",
+    "wandb_mode": "online",
+    "run_tag": "pretrained-attention-scaled-projections",
+    "smoke": false,
+    "answer_weight": 32.0,
+    "preserve_context": true,
+    "question_weight": 1.0,
+    "preserve_pretrained_attention": true
+  },
+  "checkpoint_sha256": "b7b3f9dd3e7193428bc0b3e2e444db77679031c32228a89d3521dd8d6b68bde3",
+  "sft_sha256": "7288c95445f5e553a167b1bfc9c31af9309cd14fc94828d0bd62b5d875ae438b",
+  "trainer_sha256": "67378d91ddee24e15a4ea2dab227e01b1d46f4713f86c785f3b783a1daed1145",
+  "policy_sha256": "9a971eaefca9c12215bfbef71b9c2be843bf88fd282366de36eccf8556c6869e",
+  "metrics_sha256": "2ad66edd0a66aece156570059798d1e0d4fb54f168d51304d454da249ebbe550",
+  "data_sha256": "a68479e268f8ae566e281e345435b56cc5bac8f0f235d608e1bb85b2d4a62412",
+  "split_sha256": {
+    "rl": "2322a683cc448c212520dde2f1cecc463cfbe3d47754fad3227ba5e577732077",
+    "dev": "793d5f09756607246259056e2883185f49d2710ae32311679b933ac0f09608d7",
+    "test": "6b88570762fb9858e0d4a1afb712da06a8352f768b0f68463eef7f7230db44e2"
+  },
+  "environment_sha256": "9f660ff7390d08a80fb314b35c7b36e7e9a5e426a7cea671a0644904b878a317"
+}

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/rl.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9c898d8bac3c9501e5bb047b0bdc7e570887c8f3c9b6b1db23c5d790051a66f
+size 128309045

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/sft.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5969e051254e4a2ef78db286d91b9eecebf118b6ebf6118a3b20c810977bbbb
+size 128309181

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/startup.json

@@ -0,0 +1,17 @@
+{
+  "raw_weights": "c67187cbffce452f4a1fa510a094ab550e7d804fd939c2f9bd5af68a68906290",
+  "conditioning": {
+    "recipe": "Rescale residual output projections; preserve pretrained Q/K exactly before SFT and learn them without per-step normalization.",
+    "residual_projection_scale": 0.0030092086302306023,
+    "qk_rms": null,
+    "initialized_weights": "3d52347adbfc89f12d56c6c65c221e0a546aff61cc75b4f44bdb15931edcd6a8",
+    "preserve_pretrained_attention": true
+  },
+  "parameters": 32071680,
+  "hardware": "NVIDIA A40",
+  "torch_version": "2.10.0+cu128",
+  "cuda_runtime": "12.8",
+  "torch_path": "/juice5b/u/dhei/tatsu/nano-omni/.venv/lib/python3.12/site-packages/torch/__init__.py",
+  "job": "18432567",
+  "restart": "0"
+}

cases/202f6bc88a76dd3d/projection-only-v1__32m__partial_gsm8k__pretrained-attention-scaled-projections/terminal-sft.json

@@ -0,0 +1,7 @@
+{
+  "tokens": 2097152,
+  "budget": 2097152,
+  "weights_sha256": "a37df7f3574a399036bfd9c9771a1ea1e1e5e56365bd48c9b33392c95dc75964",
+  "seconds": 202.71570450719446,
+  "role": "Only terminal full-budget SFT weights initialize RL"
+}