llm-memory-editability / docs /development-artifacts /moe-sequential-development-v1 /source /configs /moe-sequential-development-v1.json
| { | |
| "runtime": { | |
| "docker_context": "lm-memory", | |
| "docker_host": "unix:///run/docker-lm-memory/docker.sock", | |
| "image": "sha256:a994f1a1886220693eb26b7ff201147fbd4b49477d335ab2707330cab771063f", | |
| "python": "/opt/conda/bin/python", | |
| "cpus": 4, | |
| "memory": "32g", | |
| "env": { | |
| "PYTHONDONTWRITEBYTECODE": "1", | |
| "XDG_CACHE_HOME": "/tmp/cache", | |
| "TORCH_HOME": "/tmp/torch", | |
| "TORCHINDUCTOR_CACHE_DIR": "/tmp/torchinductor", | |
| "TRITON_CACHE_DIR": "/tmp/triton" | |
| } | |
| }, | |
| "batch": "moe-sequential-development-v1", | |
| "module": "moe_sequential", | |
| "gpus": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3 | |
| ], | |
| "specs": [ | |
| { | |
| "phase": "development", | |
| "model": { | |
| "vocab_size": 50257, | |
| "positions": 256, | |
| "hidden_size": 256, | |
| "attention_heads": 8, | |
| "dropout": 0.0 | |
| }, | |
| "data_file": "/ossfs/workspace/llm-memory-editability/data/sequential-transfer-v1/development.json", | |
| "data_sha256": "3f66cd319651dd4c7e5e060d2cfdb00fedf2f15839862a309097ae17b9a2d7e6", | |
| "tokenizer": "/ossfs/workspace/llm-memory-editability/data/realworld-composition-v1/tokenizer", | |
| "stage_a_steps": 8000, | |
| "stage_b_steps": 4000, | |
| "batch_size": 128, | |
| "microbatch_size": 128, | |
| "learning_rate": 0.0003, | |
| "weight_decay": 0.1, | |
| "warmup_steps": 200, | |
| "minimum_lr_fraction": 0.1, | |
| "unique_layers": 4, | |
| "repeats": 1, | |
| "evaluation_nodes": [ | |
| 0, | |
| 500, | |
| 1000, | |
| 2000, | |
| 4000, | |
| 6000, | |
| 8000, | |
| 8250, | |
| 8500, | |
| 9000, | |
| 10000, | |
| 12000 | |
| ], | |
| "evaluation_batch_size": 64, | |
| "checkpoint_interval": 500, | |
| "status_interval": 50, | |
| "cpu_threads": 2, | |
| "tracking_group": "moe-sequential-development-v1", | |
| "name": "dev-w256-m4-i870071", | |
| "history": "sequential_composition", | |
| "initialization": 870071, | |
| "sampling_seed": 871071, | |
| "dropout_seed": 872071, | |
| "arm": "M4", | |
| "initialization_reference_layers": 8, | |
| "replay_source": "full_stage_a", | |
| "architecture": "M4", | |
| "balance_coefficient": 0.01, | |
| "module_initialization": 1870074 | |
| }, | |
| { | |
| "phase": "development", | |
| "model": { | |
| "vocab_size": 50257, | |
| "positions": 256, | |
| "hidden_size": 256, | |
| "attention_heads": 8, | |
| "dropout": 0.0 | |
| }, | |
| "data_file": "/ossfs/workspace/llm-memory-editability/data/sequential-transfer-v1/development.json", | |
| "data_sha256": "3f66cd319651dd4c7e5e060d2cfdb00fedf2f15839862a309097ae17b9a2d7e6", | |
| "tokenizer": "/ossfs/workspace/llm-memory-editability/data/realworld-composition-v1/tokenizer", | |
| "stage_a_steps": 8000, | |
| "stage_b_steps": 4000, | |
| "batch_size": 128, | |
| "microbatch_size": 128, | |
| "learning_rate": 0.0003, | |
| "weight_decay": 0.1, | |
| "warmup_steps": 200, | |
| "minimum_lr_fraction": 0.1, | |
| "unique_layers": 4, | |
| "repeats": 1, | |
| "evaluation_nodes": [ | |
| 0, | |
| 500, | |
| 1000, | |
| 2000, | |
| 4000, | |
| 6000, | |
| 8000, | |
| 8250, | |
| 8500, | |
| 9000, | |
| 10000, | |
| 12000 | |
| ], | |
| "evaluation_batch_size": 64, | |
| "checkpoint_interval": 500, | |
| "status_interval": 50, | |
| "cpu_threads": 2, | |
| "tracking_group": "moe-sequential-development-v1", | |
| "name": "dev-w256-w4-i870071", | |
| "history": "sequential_composition", | |
| "initialization": 870071, | |
| "sampling_seed": 871071, | |
| "dropout_seed": 872071, | |
| "arm": "W4", | |
| "initialization_reference_layers": 8, | |
| "replay_source": "full_stage_a", | |
| "architecture": "W4", | |
| "balance_coefficient": 0.0, | |
| "module_initialization": 1870074 | |
| }, | |
| { | |
| "phase": "development", | |
| "model": { | |
| "vocab_size": 50257, | |
| "positions": 256, | |
| "hidden_size": 256, | |
| "attention_heads": 8, | |
| "dropout": 0.0 | |
| }, | |
| "data_file": "/ossfs/workspace/llm-memory-editability/data/sequential-transfer-v1/development.json", | |
| "data_sha256": "3f66cd319651dd4c7e5e060d2cfdb00fedf2f15839862a309097ae17b9a2d7e6", | |
| "tokenizer": "/ossfs/workspace/llm-memory-editability/data/realworld-composition-v1/tokenizer", | |
| "stage_a_steps": 8000, | |
| "stage_b_steps": 4000, | |
| "batch_size": 128, | |
| "microbatch_size": 128, | |
| "learning_rate": 0.0003, | |
| "weight_decay": 0.1, | |
| "warmup_steps": 200, | |
| "minimum_lr_fraction": 0.1, | |
| "unique_layers": 4, | |
| "repeats": 1, | |
| "evaluation_nodes": [ | |
| 0, | |
| 500, | |
| 1000, | |
| 2000, | |
| 4000, | |
| 6000, | |
| 8000, | |
| 8250, | |
| 8500, | |
| 9000, | |
| 10000, | |
| 12000 | |
| ], | |
| "evaluation_batch_size": 64, | |
| "checkpoint_interval": 500, | |
| "status_interval": 50, | |
| "cpu_threads": 2, | |
| "tracking_group": "moe-sequential-development-v1", | |
| "name": "dev-w256-m4-i870072", | |
| "history": "sequential_composition", | |
| "initialization": 870072, | |
| "sampling_seed": 871072, | |
| "dropout_seed": 872072, | |
| "arm": "M4", | |
| "initialization_reference_layers": 8, | |
| "replay_source": "full_stage_a", | |
| "architecture": "M4", | |
| "balance_coefficient": 0.01, | |
| "module_initialization": 1870075 | |
| }, | |
| { | |
| "phase": "development", | |
| "model": { | |
| "vocab_size": 50257, | |
| "positions": 256, | |
| "hidden_size": 256, | |
| "attention_heads": 8, | |
| "dropout": 0.0 | |
| }, | |
| "data_file": "/ossfs/workspace/llm-memory-editability/data/sequential-transfer-v1/development.json", | |
| "data_sha256": "3f66cd319651dd4c7e5e060d2cfdb00fedf2f15839862a309097ae17b9a2d7e6", | |
| "tokenizer": "/ossfs/workspace/llm-memory-editability/data/realworld-composition-v1/tokenizer", | |
| "stage_a_steps": 8000, | |
| "stage_b_steps": 4000, | |
| "batch_size": 128, | |
| "microbatch_size": 128, | |
| "learning_rate": 0.0003, | |
| "weight_decay": 0.1, | |
| "warmup_steps": 200, | |
| "minimum_lr_fraction": 0.1, | |
| "unique_layers": 4, | |
| "repeats": 1, | |
| "evaluation_nodes": [ | |
| 0, | |
| 500, | |
| 1000, | |
| 2000, | |
| 4000, | |
| 6000, | |
| 8000, | |
| 8250, | |
| 8500, | |
| 9000, | |
| 10000, | |
| 12000 | |
| ], | |
| "evaluation_batch_size": 64, | |
| "checkpoint_interval": 500, | |
| "status_interval": 50, | |
| "cpu_threads": 2, | |
| "tracking_group": "moe-sequential-development-v1", | |
| "name": "dev-w256-w4-i870072", | |
| "history": "sequential_composition", | |
| "initialization": 870072, | |
| "sampling_seed": 871072, | |
| "dropout_seed": 872072, | |
| "arm": "W4", | |
| "initialization_reference_layers": 8, | |
| "replay_source": "full_stage_a", | |
| "architecture": "W4", | |
| "balance_coefficient": 0.0, | |
| "module_initialization": 1870075 | |
| } | |
| ], | |
| "reused_runs": [ | |
| { | |
| "architecture": "D4", | |
| "initialization": 870071, | |
| "name": "dev-w256-standard-i870071", | |
| "run_dir": "/ossfs/workspace/llm-memory-editability/results/loop-kv-learning-development-v1/runs/dev-w256-standard-i870071", | |
| "files": { | |
| "run.json": "c4ffd2c629355e11883bd689c808763f066d5dadd3a315135340cfad81fa0d91", | |
| "learning.json": "dfbafbaa91569b7c4b637f60e22f611f68b9bfb7f20cc4db0250e84f59761dbb", | |
| "endpoint.json": "f8a62466533bdc183890017fef36733411facf1fc1ce7959893982672372e614", | |
| "sampling-plan.json": "d2fa1ef64e64d137e433d0e038cc72bb8e0f21d5989547b56210ab8d21a6fc00", | |
| "audit.json": "2d1309f4b9db79ee275a1371da8e846ddfa0975d13e7c795bbb4378ea734a43b", | |
| "complete.json": "4a2f336a2f81eb08b98cf7bb094260369721ad60ed85c1a9afe8e07412698949", | |
| "architecture.json": "c9a01ec71398b792066bf3941114e167dfaa55c40b07f23ea13d44daadc5a157", | |
| "stage-a.pt": "48f00373e3c5157a3a4439cdb21c300c1d9741a03e0623b0d95d422f30323fac", | |
| "latest.pt": "997ad07094049403fea3bf37e0eb382f0654c6ee90956ac75f29fda5482daf0a", | |
| "predictions-0008000.json": "650666c9e3e9833aa5174e90ea6a70bd0fc359e5e51b4bf9469f0fddb4457f4e", | |
| "predictions-0012000.json": "06fca0cb2c68fa168e042ba30a0ce0efb809c5f2a1258a11f6c31c43aa905a9f" | |
| } | |
| }, | |
| { | |
| "architecture": "D4", | |
| "initialization": 870072, | |
| "name": "dev-w256-standard-i870072", | |
| "run_dir": "/ossfs/workspace/llm-memory-editability/results/loop-kv-learning-development-v1/runs/dev-w256-standard-i870072", | |
| "files": { | |
| "run.json": "ddbedc19bb70bfa52eb0eaf577048cc133e1e65591085ccee4c94d4d747651bc", | |
| "learning.json": "b3be211db5348fbb44ae0be2ec8aae9af372a01095ebe904dad44a6ef832a0cb", | |
| "endpoint.json": "b9503dd1eee6c2702f533773d204e2c9dec71812d48c3426dd9f3cf1b022bd5e", | |
| "sampling-plan.json": "5c0ff3f3444702533d2cb7f199ed8bbf09b7b4749241d02ad0f8ade15e76a813", | |
| "audit.json": "4f352068ca0fcb96481b9c3e4250f63a5d516bdb3e9af4000ef81b40f247f46c", | |
| "complete.json": "ed55962037d233c294a8b3c815c8197845d101f4dce501c3df75c8d729627c37", | |
| "architecture.json": "bb106bc8bc6c5af5b947f1bf8fc50642f6cc87ddc3ae02dd127439db18f2681f", | |
| "stage-a.pt": "e17a1e4c1842be31c355fa108cfe148877090ac4fc85a604ef553306720ba183", | |
| "latest.pt": "6c097e710900566370553f38714f7aaa545fc37255ee5665b40910678d57d69f", | |
| "predictions-0008000.json": "7fc474a030493b01dbdafb952830a9c42169bb088762c28a635a139d01416fbc", | |
| "predictions-0012000.json": "92b104a3f0f10215b7e1f1e91389c6d39413d2453d1fb58789b20bf05790b394" | |
| } | |
| } | |
| ], | |
| "analysis": { | |
| "unit": "two initializations on one previously observed 2Wiki development split", | |
| "new_scientific_updates": 48000, | |
| "reused_updates": 24000, | |
| "architecture_search": "no new learning-rate or expert grid; validated recipe at common 3e-4", | |
| "first_layer": "fixed-panel atomic learning versus exposure, tokens and FLOPs", | |
| "second_layer": "full untrained AA at A endpoint; fixed-panel learning curves", | |
| "third_layer": "BB full endpoint and A-to-B gain; AA parent-correct per-question retention", | |
| "additional_supervision": "MoE 0.01 whole-effective-batch router balance; no extra knowledge labels", | |
| "success_interpretation": "report paired signs and costs; two initializations do not establish generality", | |
| "automatic_followup": "none" | |
| }, | |
| "additional_sources": [ | |
| "tests/test_moe_sequential.py", | |
| "scripts/report_moe_sequential.py", | |
| "scripts/audit_loop_learning_tracking.py", | |
| "docs/development-artifacts/moe-sequential-development-v1/protocol.md", | |
| "docs/development-artifacts/moe-sequential-development-v1/literature.md", | |
| "scripts/validate_moe_sequential.py", | |
| "tests/test_report_moe_sequential.py", | |
| "scripts/report_loop_learning.py" | |
| ] | |
| } | |