Archive completed gsm-visible-v2__16m__gsm8k__visible-calculation-context

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/ancestry.json

@@ -0,0 +1,9 @@
+{
+  "rl_initializer": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/gsm-visible-v2/runs/visible-calculation-context/16m/gsm8k/sft.pt",
+  "terminal_sft_weights": "0c5f77c6573a115f6e8d31e4681a79ebea114ef53a93a1ad80d9f64dc9bcc39d",
+  "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/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/archive-manifest.json

@@ -0,0 +1,37 @@
+{
+  "protocol": "54739e6657da5bce4b9b06dc3731bb846e214e551bcb1d8fcd5ffe098dfa9f89",
+  "case": {
+    "key": "gsm-visible-v2__16m__gsm8k__visible-calculation-context",
+    "folder": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/gsm-visible-v2/runs/visible-calculation-context/16m/gsm8k",
+    "reuse": false
+  },
+  "files": {
+    "sft.pt": {
+      "size": 63429146,
+      "sha256": "16257d32c83cf1c81eab1472150817518d716877adc4424e4bbbecf2c9961c99"
+    },
+    "rl.pt": {
+      "size": 63429005,
+      "sha256": "a392e226f7b50da2c453d3e39807b3c997908a94bb8578285b7cbb1e2dc5069b"
+    }
+  },
+  "complete": {
+    "task": "gsm8k",
+    "model": "16m",
+    "sft_tokens": 2097152,
+    "sft_budget": 2097152,
+    "rl_rollouts": 65536,
+    "rl_budget": 65536,
+    "terminal_sft_hash": "0c5f77c6573a115f6e8d31e4681a79ebea114ef53a93a1ad80d9f64dc9bcc39d",
+    "final_rl_hash": "2ea507509043f1a11b0f6c576f64f543a3bd72a8e53423ca088b6babe7ffa15d",
+    "sft_seconds": 355.84108114242554,
+    "rl_seconds": 826.7976912520826,
+    "no_checkpoint_selection": true,
+    "evaluation_generations": 16128,
+    "calibration_generations": 0,
+    "hint_distribution": "uniform 1..16 per query",
+    "samples_per_query": 32,
+    "development_episodes": 96,
+    "development_unique_groups": 96
+  }
+}

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/complete.json

@@ -0,0 +1,19 @@
+{
+  "task": "gsm8k",
+  "model": "16m",
+  "sft_tokens": 2097152,
+  "sft_budget": 2097152,
+  "rl_rollouts": 65536,
+  "rl_budget": 65536,
+  "terminal_sft_hash": "0c5f77c6573a115f6e8d31e4681a79ebea114ef53a93a1ad80d9f64dc9bcc39d",
+  "final_rl_hash": "2ea507509043f1a11b0f6c576f64f543a3bd72a8e53423ca088b6babe7ffa15d",
+  "sft_seconds": 355.84108114242554,
+  "rl_seconds": 826.7976912520826,
+  "no_checkpoint_selection": true,
+  "evaluation_generations": 16128,
+  "calibration_generations": 0,
+  "hint_distribution": "uniform 1..16 per query",
+  "samples_per_query": 32,
+  "development_episodes": 96,
+  "development_unique_groups": 96
+}

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/request.json

@@ -0,0 +1,67 @@
+{
+  "config": {
+    "task": "gsm8k",
+    "model": "16m",
+    "root": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/gsm-visible-v2",
+    "checkpoint": "/nlp/scr/dhei/nanogpt-ladder-v3/runs/base/16m-475529216/seed-1337/model.pt",
+    "output": "/nlp/scr/dhei/tiny-rl/experiments/refined-math-v1/iterations/gsm-visible-v2/runs/visible-calculation-context/16m/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": 0.1,
+    "qk_update_scale": 1.0,
+    "kl": 0.1,
+    "clip": 0.2,
+    "eval_queries": 96,
+    "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": "",
+    "dpo_reward": "binary",
+    "chat_stop": "none",
+    "threads": 2,
+    "seed": 2718,
+    "device": "auto",
+    "wandb_project": "tiny-rl-refined-math-v2",
+    "wandb_entity": "david-heineman",
+    "wandb_mode": "online",
+    "run_tag": "visible-calculation-context",
+    "smoke": false,
+    "evaluate_test": true,
+    "answer_weight": 32.0,
+    "preserve_context": true,
+    "question_weight": 1.0,
+    "preserve_pretrained_attention": false,
+    "rl_small_final_fraction": 0.0,
+    "stratified_evaluation": true
+  },
+  "checkpoint_sha256": "657993bbfa3d4cb1a3bfb5c6c62ac9d28d9ba968d2d72d64e41e45f5a60cb21f",
+  "sft_sha256": "ffbaa1cf024d90c13a516598a88db5ca6baf800fc3960af9ae28875b94530aa8",
+  "trainer_sha256": "846a6369ab5739a733b25d5222903b1203c008f4f06898b6ac1f4eb4e96d68cf",
+  "policy_sha256": "9a971eaefca9c12215bfbef71b9c2be843bf88fd282366de36eccf8556c6869e",
+  "metrics_sha256": "c7a912ec690a48e41f83a230c53fb280c0e4fcf9794c873b3f7eca3375f0198c",
+  "data_sha256": "2e5c286e09f78f32066940d20eceeff8b075037a33a8e880db491e0ded8ec6f7",
+  "split_sha256": {
+    "rl": "46b8fe0b4e8abb8b85701493d09fb3ed01c2b8ba0d06e3fb41a1e8b484e27f09",
+    "dev": "d28052e23b89ed568bfa21a9428e84a395cf2e8f80108e279e7429a8646c31df",
+    "test": "230a1c63808396686041d04f6e8228b1e36c6dcab6f5df3ac505a61d22ed4f0c"
+  },
+  "environment_sha256": "3a4fac01313a14b3ca9d0cb19bd014c14d23042c028cf6b20a4c3119d351d48f"
+}

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/rl.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a392e226f7b50da2c453d3e39807b3c997908a94bb8578285b7cbb1e2dc5069b
+size 63429005

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/sft.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16257d32c83cf1c81eab1472150817518d716877adc4424e4bbbecf2c9961c99
+size 63429146

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/startup.json

@@ -0,0 +1,16 @@
+{
+  "raw_weights": "efb4038c2e5aa917b1ec65814519d45ca9205a919e593cf7b7721dedb2307c18",
+  "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": 0.004151971330706384,
+    "qk_rms": 0.1,
+    "initialized_weights": "124bc4e8fe94f5d5e8b8cf4d5b994726824529fb7db82e87b28e64fdf7819deb"
+  },
+  "parameters": 15851520,
+  "hardware": "NVIDIA TITAN Xp",
+  "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": "18433561",
+  "restart": "0"
+}

cases/54739e6657da5bce/gsm-visible-v2__16m__gsm8k__visible-calculation-context/terminal-sft.json

@@ -0,0 +1,7 @@
+{
+  "tokens": 2097152,
+  "budget": 2097152,
+  "weights_sha256": "0c5f77c6573a115f6e8d31e4681a79ebea114ef53a93a1ad80d9f64dc9bcc39d",
+  "seconds": 355.84108114242554,
+  "role": "Only terminal full-budget SFT weights initialize RL"
+}