🔬Modal Research Exchange

{
  "base_model": "Qwen/Qwen3.5-4B",
  "baseline": "dimuon-lite (DiMuon with Polar Express), beta1=0.9, grad_clip_norm=0.0. Same model, rank, data order, seed, renderer, batch size, steps and LR grid. The best-LR run of each arm is compared.",
  "batch_size": 64,
  "budget": {
    "gpu_hours": 1.8,
    "max_wall_minutes": 120
  },
  "command": "python - <<'EOF'\n# Custom part: the TriMuon config exists only on the fork; the baseline uses tinker.types.DimuonOptimizerConfig() / DimuonParams.\nimport os, tinker\nfrom tinker_cookbook.recipes.chat_sl import chat_datasets  # renderer + no_robots builder\nsc = tinker.ServiceClient(base_url=os.environ[\"TINKER_BASE_URL\"])  # fork URL for the TriMuon arm\nARM = os.environ.get(\"ARM\", \"pe\"); LR = float(os.environ.get(\"LR\", \"1e-4\"))\nopt_cfg = tinker.types.DimuonOptimizerConfig() if ARM == \"pe\" else tinker.types.TrimuonOptimizerConfig()  # Trimuon* added by the fork\ntc = sc.create_lora_training_client(base_model=\"Qwen/Qwen3.5-4B\", rank=32, optimizer=opt_cfg)\ntrain, test = chat_datasets.no_robots_batches(tc.get_tokenizer(), batch_size=64, max_length=2048, seed=0)\nfor step, batch in zip(range(80), train):\n    tc.forward_backward(batch, loss_fn=\"cross_entropy\").result()\n    P = tinker.types.DimuonParams if ARM == \"pe\" else tinker.types.TrimuonParams\n    tc.optim_step(P(learning_rate=LR, beta1=0.9, grad_clip_norm=0.0)).result()\nprint(\"eval_nll\", sum(tc.forward(b, loss_fn=\"cross_entropy\").result().metrics[\"loss:mean\"] for b in test) / len(test))\ntc.save_weights_for_sampler(name=f\"rx-trimuon-{ARM}-{LR}\")\nEOF",
  "dataset": {
    "hf_path": "HuggingFaceH4/no_robots",
    "n_examples": 5120,
    "name": "no_robots",
    "split": "train (eval on test)"
  },
  "eval": {
    "benchmark": "no_robots test split",
    "metric": "mean per-token NLL on assistant tokens",
    "n_examples": 500
  },
  "hypothesis": "With a Spindle fork that adds TriMuon, LoRA SFT on Qwen3.5-4B reaches held-out no_robots loss at least 0.01 nats lower than dimuon-lite (Polar Express) at its best LR, after the same 80 steps.",
  "lora_rank": 32,
  "max_seq_len": 2048,
  "method": "Fork modal-projects/spindle and add TriMuon as an optimizer option. It keeps DiMuon's recovered momentum and preconditioning rules. It replaces only the Polar Express polar step with two outer Polar Switch passes, then four cubic Newton-Schulz refinements folded into the rank x rank multiplier. Use the FP16 polar core and TF32 disabled, as in aaron's benchmark. Deploy the fork as a throwaway copy. Train rank-32 LoRA on Qwen/Qwen3.5-4B with a short tinker SDK SFT loop on no_robots. The loop uses the cookbook renderer for the model, forward_backward with cross_entropy, then optim_step. The treatment arm runs on the fork with TriMuon. The baseline arm runs dimuon-lite on the same fork (or shared Spindle) with identical code. Each arm sweeps LR over {5e-5, 1e-4, 2e-4}: 6 runs of 80 steps each, same data order and seed. Each run evaluates held-out loss at step 80 and logs per-step optimizer wall time. Qwen3.5-4B is a scaled-down stand-in for the source's Qwen3-8B rank-32 workload. Polar Switch and AdamW arms are follow-up specs, not part of this decision.",
  "optimizer": {
    "lr": 0.0001,
    "name": "trimuon (fork)"
  },
  "steps": 80,
  "success_criterion": "Validated if best-LR TriMuon eval NLL is at least 0.01 nats below best-LR dimuon-lite eval NLL, and TriMuon's median optimizer step time is not higher. Failed if the gap is under 0.01 nats or reversed. A rerun of the best PE config with a second seed is optional, if budget allows, to check that the 0.01 margin exceeds seed noise.",
  "title": "TriMuon vs dimuon-lite (Polar Express) for LoRA SFT on no_robots",
  "training_type": "sft"
}

[
  {
    "arm": "shared_init",
    "step": 0,
    "uri": "tinker://6af90323ba7699144106782ddc682a32/weights/rx-trimuon-init-step0"
  },
  {
    "arm": "pe_lr0.0002",
    "step": 80,
    "uri": "tinker://a367e4381452827733cfa2a16b31638d/weights/rx-trimuon-pe_lr0.0002-step80"
  },
  {
    "arm": "trimuon_lr0.0002",
    "step": 80,
    "uri": "tinker://231862420b152737820c434079b61a4c/weights/rx-trimuon-trimuon_lr0.0002-step80"
  },
  {
    "arm": "pe_lr0.0032",
    "step": 80,
    "uri": "tinker://0bdf49d5a745220d8567cb4781132574/weights/rx-trimuon-pe_lr0.0032-step80"
  },
  {
    "arm": "trimuon_lr0.0032",
    "step": 80,
    "uri": "tinker://5e51405d0d0c4f35adaf7fff3e64e0b4/weights/rx-trimuon-trimuon_lr0.0032-step80"
  }
]

{
  "pe_lr0.0002": {
    "eval_nll_token_mean": 2.100793518805683
  },
  "pe_lr0.0032": {
    "eval_nll_token_mean": 1.8656645425972798
  },
  "trimuon_lr0.0002": {
    "eval_nll_token_mean": 2.1014318670873253
  },
  "trimuon_lr0.0032": {
    "eval_nll_token_mean": 1.865601469886534
  }
}

best_lr_pe

0.0032

best_lr_trimuon

0.0032

best_eval_nll_pe

1.866

best_eval_nll_trimuon

1.866

gap_pe_minus_trimuon

6.307e-05

spec_grid_best_eval_nll_pe

2.101

spec_grid_best_eval_nll_trimuon

2.101

spec_grid_best_lr_pe

0.0002

spec_grid_best_lr_trimuon

0.0002

spec_grid_gap_pe_minus_trimuon

-0.0006383

spec_grid_median_opt_step_s_pe

0.1593

spec_grid_median_opt_step_s_trimuon

0.3987

median_opt_step_s_pe_best

0.1289

median_opt_step_s_trimuon_best

0.3182

max_abs_gap_any_lr

0.001041

observed_trainer_gpu_hours_approx

1.09

• checkpoint · shared_init · step 0 · tinker://6af90323ba7699144106782ddc682a32/weights/rx-trimuon-init-step0
• checkpoint · pe_lr0.0002 · step 80 · tinker://a367e4381452827733cfa2a16b31638d/weights/rx-trimuon-pe_lr0.0002-step80
• checkpoint · trimuon_lr0.0002 · step 80 · tinker://231862420b152737820c434079b61a4c/weights/rx-trimuon-trimuon_lr0.0002-step80
• checkpoint · pe_lr0.0032 · step 80 · tinker://0bdf49d5a745220d8567cb4781132574/weights/rx-trimuon-pe_lr0.0032-step80
• checkpoint · trimuon_lr0.0032 · step 80 · tinker://5e51405d0d0c4f35adaf7fff3e64e0b4/weights/rx-trimuon-trimuon_lr0.0032-step80
• code_revision · code_revision.txt

1. Nov 4, 00:51 UTC · succeeded · done: pe_lr5e-05, pe_lr0.0001, pe_lr0.0002, pe_lr0.0004, pe_lr0.0008, pe_lr0.0016, pe_lr0.0032, trimuon_lr5e-05, trimuon_lr0.0001, trimuon_lr0.0002, trimuon_lr0.0004, trimuon_lr0.0008, trimuon_lr0.0016, trimuon_lr0.0032 · On a throwaway Spindle fork (dimuon-lite Miles branch + TriMuon reimplemented from the spec), rank-32 LoRA SFT of Qwen3.5-4B on no_robots for 80 steps gave essentially identical held-out NLL for TriM…
2. Nov 4, 00:21 UTC · running · done: pe_lr5e-05, pe_lr0.0001, pe_lr0.0002, pe_lr0.0004, trimuon_lr5e-05, trimuon_lr0.0001, trimuon_lr0.0002, trimuon_lr0.0004 · Spec grid and LR 4e-4 extension done: both arms keep improving at the top LR (4e-4: PE 1.95142 vs TriMuon 1.95246), TriMuon's median optimizer step is ~2.4x slower. Running a second symmetric extensi…
3. Nov 4, 00:11 UTC · running · done: pe_lr5e-05, pe_lr0.0001, pe_lr0.0002, trimuon_lr5e-05, trimuon_lr0.0001, trimuon_lr0.0002 · All 6 spec runs (PE and TriMuon x LR {5e-5,1e-4,2e-4}, 80 steps) finished on the fork; both arms select LR 2e-4 (grid edge), where TriMuon eval NLL 2.10143 vs PE 2.10079 and TriMuon's median optimize…
4. Nov 3, 23:47 UTC · running · Fork deployed; preflight probe passed for both arms (server reports optimizer_family_code 1=dimuon/PE and 2=trimuon after weight restore); local CPU parity: PE path bit-exact vs 228b59d, TriMuon vs F…
5. Nov 3, 23:44 UTC · running · Spindle fork spindle-fork-run_9a86933e38 deployed (dimuon-lite Miles branch + TriMuon reimplemented from spec); preflight probe running.

adaptation_manifests

[{'arm': 'pe_lr0.0002', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': ['s1:45c351', 's2:8a6f13', 's3:191621', 's4:38725d', 's5:8aed40', 's6:6ac1b0', 's7:b344e1', 's8:37caa3', 's9:b433b5', 's10:d88cd8', 's11:2c0453', 's12:107ca8', 's13:f0d0fe', 's14:472b0f', 's15:7105be', 's16:425716', 's17:c5602a', 's18:701dc4', 's19:1dd4c1', 's20:f9c88c', 's21:0a487c', 's22:f6c667', 's23:cc03e7', 's24:e02a09', 's25:c5eb24', 's26:ef92d7', 's27:0a389b', 's28:35f385', 's29:de8035', 's30:e20a2c', 's31:d5ce1c', 's32:a0d22e', 's33:aa1406', 's34:68d2e9', 's35:40b4b0', 's36:c297b1', 's37:4c1b74', 's38:f9f32e', 's39:05a8e9', 's40:f91b8f', 's41:f1c136', 's42:9965cb', 's43:de610c', 's44:85f9df', 's45:698a22', 's46:3681a0', 's47:177de4', 's48:8b1ea4', 's49:bbd283', 's50:1dd8d9', 's51:0861f7', 's52:495689', 's53:d3ceb2', 's54:5ddbe8', 's55:38f74d', 's56:aaf09d', 's57:d1db0d', 's58:ed5839', 's59:8f0ea7', 's60:757b50', 's61:55bd9d', 's62:2ef2ef', 's63:46e35e', 's64:9f96b4', 's65:d10ca7', 's66:898648', 's67:62cfb0', 's68:2a84c3', 's69:2c5108', 's70:c8c508', 's71:0f3538', 's72:1d6f2d', 's73:1228cb', 's74:80f675', 's75:19c5ff', 's76:e065f9', 's77:42fec6', 's78:f62f99', 's79:57fd27', 's80:ecefbc'], 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd', 'id_format': 's{step}:{first 6 hex of sha256(comma-joined no_robots prompt_ids consumed)}'}, {'arm': 'pe_lr5e-05', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'pe_lr0.0001', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'pe_lr0.0004', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'pe_lr0.0008', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'pe_lr0.0016', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'pe_lr0.0032', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr5e-05', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0001', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0002', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0004', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0008', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0016', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}, {'arm': 'trimuon_lr0.0032', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': 'identical to pe_lr0.0002', 'adaptation_batch_hash': 'd88b8aa6b7994c28dbf609dc40fa29d34b2d5184f493d6f0b8edf6777e5ae3fd'}]

correctness_preflight

{'required': True, 'status_note': 'best-effort per run instructions (spec has no tolerances): local CPU parity of the exact fork optimizer code', 'paths': [{'id': 'trimuon_optimizer', 'fixed_input': {'sha256': 'e5cc8cee9a49f85ba80685605348f47439d1da3350b86bc851902dd7e720d99e', 'description': 'correctness_fixture.pt: seeded (1234) LoRA-like factors A 32x2560, B 2560x32, fused-qkv B 576x32 (row split 128/32/32), anisotropic grads, BF16 previous momentum (warm) or zero (fresh); optimizer reset per probe; rank 32'}, 'baseline_revision': 'modal-projects/miles 228b59d90 (dimuon-lite = parent of withdrawal 45937c3b)', 'reference_revision': 'independent FP64 reference in correctness_preflight.py (PS2-55 from polar-switch@6f382f1 applied unfolded + 4 explicit cubic NS)', 'implementation_revision': 'miles fork 0ab8c3c178a1631fbcbaab5e84804af487bdefb2', 'precision': 'CPU; polar core FP16 matmuls with FP32 normalization/shift, TF32 disabled; FP32 momentum work, BF16 stored momentum', 'normalization': 'direction Frobenius-normalized to sqrt(rank) per block; lr*(rank/32)^0.175; update = p_new - p_old', 'errors': {'baseline_pe_update_warm': {'max_abs_error': 0.0, 'reference_max_abs': 5.417875945568085e-05, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000002}, 'baseline_pe_update_fresh': {'max_abs_error': 0.0, 'reference_max_abs': 4.151090979576111e-05, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0000000000000002}, 'reference_trimuon_update_warm': {'max_abs_error': 1.9760320646345705e-07, 'reference_max_abs': 5.062433262543459e-05, 'relative_l2_error': 0.002398292411463761, 'cosine_similarity': 0.9999971240992352}, 'reference_trimuon_update_fresh': {'max_abs_error': 1.1943913745355486e-07, 'reference_max_abs': 3.7186884629232666e-05, 'relative_l2_error': 0.001867388510372267, 'cosine_similarity': 0.9999982564310306}}, 'lr_scaling': [{'learning_rate': 5e-05, 'update_norm': 0.000632455805392149}, {'learning_rate': 0.0001, 'update_norm': 0.001264912027666012}, {'learning_rate': 0.0002, 'update_norm': 0.0025298222735317074}]}]}

correctness_validation

{'status': 'unverified', 'required': True, 'checks': [], 'reason': 'Correctness preflight was not recorded.', 'evidence_source': 'executor report; remote execution is not attested', 'applicability': 'unknown', 'reported_verdict': 'inconclusive'}

evaluation

{'version': 1, 'evaluators': {'eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb': {'code_sha256': 'c96a678983407dca2b34b062b5df4de329df4a6c30a08a772d2def4c9c0bcb4f', 'prompt_template': "tinker_cookbook renderer qwen3_5 (model_info.get_recommended_renderer_name('Qwen/Qwen3.5-4B')), conversation_to_datum(messages, renderer, max_length=2048, TrainOnWhat.ALL_ASSISTANT_MESSAGES); full no_robots test conversation teacher-forced", 'decoding': {'temperature': 0, 'max_tokens': 0, 'stop': [], 'mode': 'teacher_forced_forward_cross_entropy'}, 'parser': "loss_eval.py LossEval('token_mean'): -sum(logprobs where weight>0)/count over all items", 'items': {'manifest_ref': 'eval_items.jsonl', 'manifest_sha256': '9e8df8da44421c2fbb678a81672ad9b15ea25dd7c2f5ab53871caf7e258023aa'}}}, 'scores': [{'arm': 'pe_lr0.0002', 'metric': 'eval_nll_token_mean', 'value': 2.100793518805683, 'evaluator_id': 'eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}, {'arm': 'trimuon_lr0.0002', 'metric': 'eval_nll_token_mean', 'value': 2.1014318670873253, 'evaluator_id': 'eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}, {'arm': 'pe_lr0.0032', 'metric': 'eval_nll_token_mean', 'value': 1.8656645425972798, 'evaluator_id': 'eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}, {'arm': 'trimuon_lr0.0032', 'metric': 'eval_nll_token_mean', 'value': 1.865601469886534, 'evaluator_id': 'eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb', 'seed': 0, 'predictions_ref': 'predictions.jsonl'}], 'scores_note': 'all 14 runs scored with the same evaluator; values equal loss_eval.arms', 'comparisons': [{'metric': 'eval_nll_token_mean', 'arms': ['pe_lr0.0002', 'trimuon_lr0.0002']}, {'metric': 'eval_nll_token_mean', 'arms': ['pe_lr0.0032', 'trimuon_lr0.0032']}]}

evaluator_validation

{'status': 'comparable', 'issues': [], 'comparisons': [{'metric': 'eval_nll_token_mean', 'status': 'comparable', 'arms': ['pe_lr0.0002', 'pe_lr0.0032', 'trimuon_lr0.0002', 'trimuon_lr0.0032'], 'evaluator_ids': ['eec79d515f9342bf4bf919902fabe9e6665013579f02aff8db4b962009bdaacb'], 'missing_arms': []}]}

feasibility

{'status': 'unknown', 'problems': ['Re-projection required: lora_rank changed or was not recorded', 'Re-projection required: dataset changed or was not recorded', 'Re-projection required: steps changed or was not recorded', 'Re-projection required: batch_size changed or was not recorded', 'Re-projection required: max_seq_len changed or was not recorded', 'Re-projection required: arms changed or was not recorded', 'Re-projection required: n_examples changed or was not recorded', 'Re-projection required: n_eval changed or was not recorded', 'Primary endpoint changed or was not recorded'], 'wall_seconds': None, 'gpu_hours': None, 'known_gpu_hours': 1.0653148148148148, 'phases': {'setup': {'wall_seconds': 505.0, 'planned_units': 1, 'planned_samples': 0.0, 'gpu_hours': 0.14027777777777778}, 'training': {'wall_seconds': 2893.3333333333335, 'planned_units': 1120, 'planned_samples': 71680.0, 'gpu_hours': 0.8037037037037037}, 'evaluation': {'wall_seconds': 436.79999999999995, 'planned_units': 7000, 'planned_samples': 7000.0, 'gpu_hours': 0.12133333333333332}, 'recovery': {'wall_seconds': 1200, 'gpu_hours': None}}, 'generation_counts': None, 'generation_applicability': {'check': 'generation_audit', 'status': 'unknown', 'enforce': False, 'reasons': ['spec has no experiment_features: generated_training_data'], 'features': {'generated_training_data': None}}, 'record': {'config': {'base_model': 'Qwen/Qwen3.5-4B', 'probe': '1 example, 1 optim_step per arm', 'trainer': '1x H100, 8 LoRA slots'}, 'effective_config': {'base_model': 'Qwen/Qwen3.5-4B', 'arms': ['pe', 'trimuon'], 'lr_grid': [5e-05, 0.0001, 0.0002, 0.0004, 0.0008, 0.0016, 0.0032], 'steps': 80, 'batch_size': 64, 'max_seq_len': 2048, 'n_eval': 500}, 'primary_endpoint': 'mean per-token NLL on assistant tokens (no_robots test, 500)', 'source': 'probe_pe.log, probe_tri.log, runs/*.jsonl', 'measurements': {'setup': {'seconds': 505, 'units': 1, 'planned_units': 1, 'samples': 0, 'gpu_count': 1, 'allocation_source': 'fork config trainer_gpus_per_node=1'}, 'training': {'seconds': 1240, 'units': 480, 'planned_units': 1120, 'samples': 30720, 'gpu_count': 1, 'allocation_source': 'fork config trainer_gpus_per_node=1'}, 'evaluation': {'seconds': 31.2, 'units': 500, 'planned_units': 7000, 'samples': 500, 'gpu_count': 1, 'allocation_source': 'fork config trainer_gpus_per_node=1'}}, 'generation': None, 'recovery': {'seconds': 1200, 'gpu_count': None, 'allocation_source': None}}, 'recorded_at': 1793753476.9242814}

feasibility_required

True

initialization_provenance

{'version': 1, 'shared_init_checkpoint': 'tinker://6af90323ba7699144106782ddc682a32/weights/rx-trimuon-init-step0', 'adapter_sha256_all_arms': 'af593f8bef000d77412d58da0522e0c2e6cdd638e9471052d76ac0e9750d454c', 'arms': {'pe_lr0.0002': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'adapter_sha256': 'af593f8bef000d77412d58da0522e0c2e6cdd638e9471052d76ac0e9750d454c', 'source': 'local_hash', 'evidence_ref': 'sha256 of init miles/__0_0.distcp (modal volume get) + load_state ok in runs/pe_lr0.0002.jsonl', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 0, 'before_next_update': True, 'requested': 'dimuon', 'actual_family': 'dimuon', 'actual_config': {'type': 'dimuon', 'version': 1}, 'source': 'server_response', 'evidence_ref': 'get_info after create, runs/pe_lr0.0002.jsonl'}, {'after_restore': 1, 'before_next_update': True, 'requested': 'dimuon', 'actual_family': 'dimuon', 'actual_config': {'type': 'dimuon', 'version': 1, 'family_code': 1.0, 'polar_express': 1.0}, 'source': 'server_response', 'evidence_ref': 'get_info after load_state + optim_step optimizer/* metrics, runs/pe_lr0.0002.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '0441a4d96eb4ab90fc8f2a9d6c545694f67ae1e2fafd5d468179e493ac651948', 'output_sha256': '6ae994bfc7012c81bff0f0644ba33ef178d185f925e069100f54ca27049df796', 'evidence_ref': 'runs/pe_lr0.0002.jsonl', 'before_first_update': True}}, 'trimuon_lr0.0002': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'adapter_sha256': 'af593f8bef000d77412d58da0522e0c2e6cdd638e9471052d76ac0e9750d454c', 'source': 'local_hash', 'evidence_ref': 'sha256 of init miles/__0_0.distcp + load_state ok in runs/trimuon_lr0.0002.jsonl', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 0, 'before_next_update': True, 'requested': 'trimuon', 'actual_family': 'trimuon', 'actual_config': {'type': 'trimuon', 'version': 1}, 'source': 'server_response', 'evidence_ref': 'get_info after create, runs/trimuon_lr0.0002.jsonl'}, {'after_restore': 1, 'before_next_update': True, 'requested': 'trimuon', 'actual_family': 'trimuon', 'actual_config': {'type': 'trimuon', 'version': 1, 'family_code': 2.0, 'polar_trimuon': 1.0}, 'source': 'server_response', 'evidence_ref': 'get_info after load_state + optim_step optimizer/* metrics, runs/trimuon_lr0.0002.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '0441a4d96eb4ab90fc8f2a9d6c545694f67ae1e2fafd5d468179e493ac651948', 'output_sha256': '6ae994bfc7012c81bff0f0644ba33ef178d185f925e069100f54ca27049df796', 'evidence_ref': 'runs/trimuon_lr0.0002.jsonl', 'before_first_update': True}}, 'pe_lr0.0032': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'adapter_sha256': 'af593f8bef000d77412d58da0522e0c2e6cdd638e9471052d76ac0e9750d454c', 'source': 'local_hash', 'evidence_ref': 'runs/pe_lr0.0032.jsonl', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 0, 'before_next_update': True, 'requested': 'dimuon', 'actual_family': 'dimuon', 'actual_config': {'type': 'dimuon', 'version': 1}, 'source': 'server_response', 'evidence_ref': 'runs/pe_lr0.0032.jsonl'}, {'after_restore': 1, 'before_next_update': True, 'requested': 'dimuon', 'actual_family': 'dimuon', 'actual_config': {'type': 'dimuon', 'version': 1, 'family_code': 1.0}, 'source': 'server_response', 'evidence_ref': 'runs/pe_lr0.0032.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '0441a4d96eb4ab90fc8f2a9d6c545694f67ae1e2fafd5d468179e493ac651948', 'output_sha256': '5c2ff31e9d93bc80241c90cbe467ddc7e7dae1326f2ad9b22d8c2845d412207b', 'evidence_ref': 'runs/pe_lr0.0032.jsonl', 'before_first_update': True}}, 'trimuon_lr0.0032': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'adapter_sha256': 'af593f8bef000d77412d58da0522e0c2e6cdd638e9471052d76ac0e9750d454c', 'source': 'local_hash', 'evidence_ref': 'runs/trimuon_lr0.0032.jsonl', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 0, 'before_next_update': True, 'requested': 'trimuon', 'actual_family': 'trimuon', 'actual_config': {'type': 'trimuon', 'version': 1}, 'source': 'server_response', 'evidence_ref': 'runs/trimuon_lr0.0032.jsonl'}, {'after_restore': 1, 'before_next_update': True, 'requested': 'trimuon', 'actual_family': 'trimuon', 'actual_config': {'type': 'trimuon', 'version': 1, 'family_code': 2.0}, 'source': 'server_response', 'evidence_ref': 'runs/trimuon_lr0.0032.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '0441a4d96eb4ab90fc8f2a9d6c545694f67ae1e2fafd5d468179e493ac651948', 'output_sha256': 'b64ea158cd93ad1b520590d5e8b090ed9bc09670d1ea815778374d1f1e3b5a3e', 'evidence_ref': 'runs/trimuon_lr0.0032.jsonl', 'before_first_update': True}}, 'other_10_runs': {'summary': 'same shared init checkpoint and adapter_sha256, restore_count=1, server-observed family matches requested arm on every optim_step; initial_output_sha256 = 6ae994bf... for all except pe_lr0.0004 (b64ea158...), trimuon_lr0.0004 (5c2ff31e...), trimuon_lr0.0016 (7f9553f9...)'}}, 'note': 'All runs load weights only (restore_optimizer=False) from one step-0 checkpoint saved by a fresh dimuon client. Initial-output hashes match for 10 of 14 runs; the differing ones ran under different concurrent packing and the same hash values recur across pe and trimuon runs, so the difference is attributed to kernel/packing nondeterminism, not init (not proven). The concrete family/polar flags come from optim_step responses (reported with update 1, not strictly before it). No mid-training restore occurs, so the create-save-restore fixed-gradient check was not run.'}

launch

{'experiment_id': 'exp_ff293ca443', 'context_status': 'legacy_unknown', 'superseded': False, 'mode': 'redo', 'purpose': "Repeat the selected saved experiment version to check reproducibility. This repetition tests that version's hypothesis and methods."}

progress_updated_at

1793753456.7063627

reported_outcome_reason

insufficient_evidence

reported_verdict

inconclusive

restoration_validation

{'status': 'unverified', 'required': False, 'checks': [], 'reason': 'Initialization and optimizer evidence was not required by this saved spec.', 'applicability': 'unknown'}

started_at

1793748389.8436682