Aggressive autoresearch/20260323 083105

.autoresearch/session.json

@@ -1,23 +1,23 @@
 {
-  "accepted_artifact_bytes": 7770766,
-  "accepted_commit": "09535e0bf240ac86aa853ded331f797fe1560688",
-  "accepted_commit_short": "09535e0",
-  "accepted_results_path": "runs/autoresearch_5090/runs/ar5090-20260321-143929/results.json",
-  "accepted_run_id": "ar5090-20260321-143929",
-  "accepted_val_bpb": 1.5584812964313723,
+  "accepted_artifact_bytes": 15732516,
+  "accepted_commit": "bd5e987a00b69a9285e97534d376b5cad759e4d6",
+  "accepted_commit_short": "bd5e987",
+  "accepted_results_path": "runs/autoresearch_5090/runs/ar5090-20260323-074105/results.json",
+  "accepted_run_id": "ar5090-20260323-074105",
+  "accepted_val_bpb": 1.5254880853616,
   "baseline_artifact_bytes": 7174403,
   "baseline_results_path": "/workspace/autoresearch-parameter-golf/runs/autoresearch_5090/runs/baseline_5090_5min_rerun/results.json",
   "baseline_run_id": "baseline_5090_5min_rerun",
   "baseline_val_bpb": 1.5701376649442806,
   "created_at_unix": 1774089643.238404,
-  "current_branch": "autoresearch/20260319-181724",
+  "current_branch": "aggressive-autoresearch/20260323-083105",
   "current_experiment": null,
-  "latest_artifact_bytes": 7911607,
+  "latest_artifact_bytes": 15752024,
   "latest_decision": "reverted",
-  "latest_results_path": "/workspace/autoresearch-parameter-golf/runs/autoresearch_5090/runs/ar5090-20260321-154606/results.json",
-  "latest_run_id": "ar5090-20260321-154606",
+  "latest_results_path": "/workspace/autoresearch-parameter-golf/runs/autoresearch_5090/runs/ar5090-20260323-100724/results.json",
+  "latest_run_id": "ar5090-20260323-100724",
   "latest_status": "success",
-  "latest_val_bpb": 1.5672447217875,
+  "latest_val_bpb": 1.5398915807983933,
   "repo_root": "/workspace/autoresearch-parameter-golf",
   "schema_version": "pgolf.autoresearch_session.v1",
   "search_policy": {
@@ -58,5 +58,5 @@
   },
   "state_dir": "/workspace/autoresearch-parameter-golf/.autoresearch",
   "status": "ready",
-  "updated_at_unix": 1774108535.5078502
+  "updated_at_unix": 1774260881.4368758
 }

COMPETITIVE_PRIORS.md

@@ -216,3 +216,15 @@ Treat the search as a story board, not as free-form mutation roulette:
 - If recent history is narrow, switch stories before refining.
 - If a story requires a missing self-contained module in `train.py`, implement the module instead of downgrading the story to a precision micro-tune.
 - Keep module-writing stories isolated so the result is interpretable and easy to revert.
+
+## Hot Steering Override
+For the next search block, prioritize macro-topology diversity over local precision refinements.
+
+Required macro axes:
+- shared_layers vs recurrence_loops
+- tail_layers
+- d_model
+- seq_len / curriculum
+- local-module family
+
+Do not spend more than two additional runs inside the same outer carrier topology without changing one of the axes above.

configs/promoted/autoresearch_5090_best.json

@@ -0,0 +1,125 @@
+{
+  "artifact_bundle_name": "submission_bundle",
+  "benchmark_eval_repeats": 1,
+  "benchmark_only": false,
+  "benchmark_train_steps": 3,
+  "checkpoint_every": 0,
+  "counted_code_paths": [
+    "train.py"
+  ],
+  "deterministic": true,
+  "eval_first_step": false,
+  "evaluate_only": false,
+  "grad_accum_steps": 2,
+  "iterations": 200000,
+  "lawa_last_n_steps": 100,
+  "load_artifact_path": null,
+  "log_every": 10,
+  "max_wallclock_seconds": 300.0,
+  "metrics_jsonl_path": null,
+  "model": {
+    "adapter_alpha": 16.0,
+    "adapter_rank": 8,
+    "adapter_targets": [
+      "q",
+      "k",
+      "v",
+      "attn_out",
+      "mlp_in",
+      "mlp_out"
+    ],
+    "attn_dropout": 0.0,
+    "attn_fake_quant_during_train": false,
+    "d_model": 512,
+    "emb_init_std": 0.02,
+    "fake_quant_during_train": true,
+    "fake_quant_start_step": 160,
+    "final_tail_mlp_fake_quant_during_train": false,
+    "final_tail_neighbor_mixer": true,
+    "final_tail_q_low_rank": 0,
+    "final_tail_smear_gate": false,
+    "logit_softcap": 30.0,
+    "mlp_mult": 2,
+    "non_recurrent_mlp_hidden_bonus": 3584,
+    "num_heads": 8,
+    "num_kv_heads": 4,
+    "penultimate_tail_mlp_fake_quant_during_train": false,
+    "q_low_rank": 128,
+    "qk_gain_init": 1.0,
+    "recurrence_loops": 0,
+    "resid_dropout": 0.0,
+    "rope_base": 10000.0,
+    "seq_len": 768,
+    "shared_layers": 0,
+    "shared_mlp_fake_quant_during_train": null,
+    "shared_mlp_hidden_bonus": 0,
+    "shared_q_low_rank": null,
+    "stem_layers": 0,
+    "tail_layers": 3,
+    "tail_mlp_hidden_bonuses": [
+      3840,
+      3584,
+      3328
+    ],
+    "tie_embeddings": true,
+    "vocab_size": 1024
+  },
+  "optim": {
+    "adam_eps": 1e-08,
+    "beta1": 0.9,
+    "beta2": 0.95,
+    "embed_lr": 0.003,
+    "grad_clip_norm": 1.0,
+    "head_lr": 0.003,
+    "matrix_lr": 0.012,
+    "min_lr_scale": 0.1,
+    "muon_backend_steps": 5,
+    "muon_momentum": 0.95,
+    "scalar_lr": 0.003,
+    "warmdown_steps": 80,
+    "warmup_steps": 20,
+    "weight_decay": 0.0
+  },
+  "output_dir": "./runs/autoresearch_5090/promoted_current",
+  "quant": {
+    "clip_percentile": 96.5,
+    "keep_float_max_numel": 65536,
+    "keep_float_name_patterns": [
+      "norm",
+      "scale",
+      "gain",
+      "adapter",
+      "lm_head",
+      "tok_emb.weight",
+      "tail.2.mlp.",
+      "tail.2.attn.q_proj.weight",
+      "tail.2.attn.out_proj.weight"
+    ],
+    "keep_float_store_dtype": "torch.float16",
+    "low_bit_bits": 6,
+    "low_bit_name_patterns": [
+      "mlp.fc.weight",
+      "mlp.proj.weight"
+    ],
+    "scale_store_dtype": "torch.float16",
+    "zlib_level": 9
+  },
+  "results_tsv_path": "./runs/autoresearch_5090/results.tsv",
+  "resume_from": null,
+  "run_name": null,
+  "save_final_quantized": true,
+  "seed": 1337,
+  "tensorboard_log_dir": null,
+  "tokenizer_path": "./data/tokenizers/fineweb_1024_bpe.model",
+  "train_batch_tokens": 61440,
+  "train_pattern": "./data/datasets/fineweb10B_sp1024/fineweb_train_*.bin",
+  "train_phase_only": false,
+  "train_seq_len_min": 640,
+  "train_seq_len_warmup_steps": 160,
+  "use_compile": false,
+  "use_lawa": false,
+  "val_batch_tokens": 122880,
+  "val_every": 0,
+  "val_pattern": "./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin",
+  "verify_export_reload": false
+}

configs/promoted/autoresearch_h100_1x_best.json

@@ -0,0 +1,125 @@
+{
+  "artifact_bundle_name": "submission_bundle",
+  "benchmark_eval_repeats": 1,
+  "benchmark_only": false,
+  "benchmark_train_steps": 3,
+  "checkpoint_every": 100,
+  "counted_code_paths": [
+    "train.py"
+  ],
+  "deterministic": true,
+  "eval_first_step": false,
+  "evaluate_only": false,
+  "grad_accum_steps": 2,
+  "iterations": 200000,
+  "lawa_last_n_steps": 100,
+  "load_artifact_path": null,
+  "log_every": 20,
+  "max_wallclock_seconds": 600.0,
+  "metrics_jsonl_path": null,
+  "model": {
+    "adapter_alpha": 16.0,
+    "adapter_rank": 8,
+    "adapter_targets": [
+      "q",
+      "k",
+      "v",
+      "attn_out",
+      "mlp_in",
+      "mlp_out"
+    ],
+    "attn_dropout": 0.0,
+    "attn_fake_quant_during_train": false,
+    "d_model": 512,
+    "emb_init_std": 0.02,
+    "fake_quant_during_train": true,
+    "fake_quant_start_step": 160,
+    "final_tail_mlp_fake_quant_during_train": false,
+    "final_tail_neighbor_mixer": true,
+    "final_tail_q_low_rank": 0,
+    "final_tail_smear_gate": false,
+    "logit_softcap": 30.0,
+    "mlp_mult": 2,
+    "non_recurrent_mlp_hidden_bonus": 3584,
+    "num_heads": 8,
+    "num_kv_heads": 4,
+    "penultimate_tail_mlp_fake_quant_during_train": false,
+    "q_low_rank": 128,
+    "qk_gain_init": 1.0,
+    "recurrence_loops": 0,
+    "resid_dropout": 0.0,
+    "rope_base": 10000.0,
+    "seq_len": 768,
+    "shared_layers": 0,
+    "shared_mlp_fake_quant_during_train": null,
+    "shared_mlp_hidden_bonus": 0,
+    "shared_q_low_rank": null,
+    "stem_layers": 0,
+    "tail_layers": 3,
+    "tail_mlp_hidden_bonuses": [
+      3840,
+      3584,
+      3328
+    ],
+    "tie_embeddings": true,
+    "vocab_size": 1024
+  },
+  "optim": {
+    "adam_eps": 1e-08,
+    "beta1": 0.9,
+    "beta2": 0.95,
+    "embed_lr": 0.003,
+    "grad_clip_norm": 1.0,
+    "head_lr": 0.003,
+    "matrix_lr": 0.012,
+    "min_lr_scale": 0.1,
+    "muon_backend_steps": 5,
+    "muon_momentum": 0.95,
+    "scalar_lr": 0.003,
+    "warmdown_steps": 80,
+    "warmup_steps": 20,
+    "weight_decay": 0.0
+  },
+  "output_dir": "./runs/runpod_h100_1x_10min/promoted",
+  "quant": {
+    "clip_percentile": 96.5,
+    "keep_float_max_numel": 65536,
+    "keep_float_name_patterns": [
+      "norm",
+      "scale",
+      "gain",
+      "adapter",
+      "lm_head",
+      "tok_emb.weight",
+      "tail.2.mlp.",
+      "tail.2.attn.q_proj.weight",
+      "tail.2.attn.out_proj.weight"
+    ],
+    "keep_float_store_dtype": "torch.float16",
+    "low_bit_bits": 6,
+    "low_bit_name_patterns": [
+      "mlp.fc.weight",
+      "mlp.proj.weight"
+    ],
+    "scale_store_dtype": "torch.float16",
+    "zlib_level": 9
+  },
+  "results_tsv_path": "./runs/runpod_h100_1x_10min/results.tsv",
+  "resume_from": null,
+  "run_name": null,
+  "save_final_quantized": true,
+  "seed": 1337,
+  "tensorboard_log_dir": null,
+  "tokenizer_path": "./data/tokenizers/fineweb_1024_bpe.model",
+  "train_batch_tokens": 61440,
+  "train_pattern": "./data/datasets/fineweb10B_sp1024/fineweb_train_*.bin",
+  "train_phase_only": false,
+  "train_seq_len_min": 640,
+  "train_seq_len_warmup_steps": 160,
+  "use_compile": false,
+  "use_lawa": true,
+  "val_batch_tokens": 122880,
+  "val_every": 100,
+  "val_pattern": "./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin",
+  "verify_export_reload": true
+}