{ "completed_steps": 8000, "config": { "arch": "geometric", "context": 384, "heads": 24, "mlp_hidden": 675, "pattern": "rrarrarrarrarrar", "pointer": { "dim": 32, "init_seed": 1 }, "read": "l2", "rotation": true, "seed": 1, "vocab_size": 4096, "width": 1536 }, "curve": [ { "dev_first_four_nll": 0.2577194934592274, "dev_pointer_hit_rate": 0.18042813455657492, "dev_pointer_mean_gate": 0.009520117232463415, "dev_pointer_reachable_rate": 0.5474006116207951, "dev_response_nll": 0.10950227250492943, "grad_norm": 3.673137950812722, "lr": 0.0001330424996255061, "step": 2000, "supervised_target_visits": 1280724, "train_response_nll": 1.0088697372674942, "train_seconds": 326.25358548500054 }, { "dev_first_four_nll": 0.22705570096026126, "dev_pointer_hit_rate": 0.27217125382262997, "dev_pointer_mean_gate": 0.00872693953621794, "dev_pointer_reachable_rate": 0.5474006116207951, "dev_response_nll": 0.09281377307786252, "grad_norm": 2.5930435873267474, "lr": 0.00008524511554710438, "step": 4000, "supervised_target_visits": 2576779, "train_response_nll": 0.8766637990437448, "train_seconds": 654.2726661280012 }, { "dev_first_four_nll": 0.2650291993524254, "dev_pointer_hit_rate": 0.25688073394495414, "dev_pointer_mean_gate": 0.011431238227270034, "dev_pointer_reachable_rate": 0.5474006116207951, "dev_response_nll": 0.10511637535111114, "grad_norm": 2.8143733249551732, "lr": 0.00003576071715984797, "step": 6000, "supervised_target_visits": 3876794, "train_response_nll": 0.7702486326508224, "train_seconds": 982.3928720550002 }, { "dev_first_four_nll": 0.2315104865340856, "dev_pointer_hit_rate": 0.26605504587155965, "dev_pointer_mean_gate": 0.011816133449776159, "dev_pointer_reachable_rate": 0.5474006116207951, "dev_response_nll": 0.09195867602596103, "grad_norm": 3.0009986486841838, "lr": 0.000015000005475002362, "step": 8000, "supervised_target_visits": 5155535, "train_response_nll": 0.6823601159602404, "train_seconds": 1308.3076496690032 } ], "development_selection": "uor-r4.dialogue-development/source-sha256-rank-v1", "executable": { "bytes": 34788352, "path": "/target-b/release/examples/geometric-stack", "sha256": "628c854d72a545a54906d72574c319407b399721df36333c7d52a7ab18f0586b" }, "final_development": { "conditioning": "full_original_prefix", "eos_targets": 32, "first_four_response_targets_mean_nll": 0.2315104865340856, "first_four_targets": 128, "per_source": [ { "eligible_responses": 1223, "eos_targets": 32, "first_four_response_targets_mean_nll": 0.2315104865340856, "first_four_targets": 128, "label": "m-world-v2.train", "pointer": { "definitions": "over the scored targets: mean_gate is the mean gate g_t; pointer_hit_rate is the fraction whose most attended source (lowest position on a tie) holds the target token; target_reachable_rate is the fraction whose target token is held by any source with attention (the hit rate's ceiling); mean_copy_mass is the mean p_copy(target) before the gate; gate_bce is the mean of -log g_t on reachable targets and -log(1 - g_t) on the others, and pointer_nll the sum of -log p_copy(target) over reachable targets divided by all scored targets (the two copy-gate supervision terms, reported whether or not a run trains them)", "gate_bce": 5.255796121810704, "mean_copy_mass": 0.21853918230027355, "mean_gate": 0.011816133449776159, "pointer_hit_rate": 0.26605504587155965, "pointer_nll": 1.2457404604805489, "scored_targets": 327, "target_reachable_rate": 0.5474006116207951 }, "response_ids": [ 962, 1138, 101, 1102, 613, 671, 283, 1185, 421, 484, 1129, 639, 170, 1063, 677, 200, 143, 714, 424, 928, 880, 1015, 740, 999, 427, 420, 395, 76, 1221, 936, 1104, 673 ], "response_mean_nll": 0.09195867602596103, "selected_responses": 32, "source_index": 0, "supervised_targets": 327 } ], "pointer": { "definitions": "over the scored targets: mean_gate is the mean gate g_t; pointer_hit_rate is the fraction whose most attended source (lowest position on a tie) holds the target token; target_reachable_rate is the fraction whose target token is held by any source with attention (the hit rate's ceiling); mean_copy_mass is the mean p_copy(target) before the gate; gate_bce is the mean of -log g_t on reachable targets and -log(1 - g_t) on the others, and pointer_nll the sum of -log p_copy(target) over reachable targets divided by all scored targets (the two copy-gate supervision terms, reported whether or not a run trains them)", "gate_bce": 5.255796121810704, "mean_copy_mass": 0.21853918230027355, "mean_gate": 0.011816133449776159, "pointer_hit_rate": 0.26605504587155965, "pointer_nll": 1.2457404604805489, "scored_targets": 327, "target_reachable_rate": 0.5474006116207951 }, "response_ids": [ 962, 1138, 101, 1102, 613, 671, 283, 1185, 421, 484, 1129, 639, 170, 1063, 677, 200, 143, 714, 424, 928, 880, 1015, 740, 999, 427, 420, 395, 76, 1221, 936, 1104, 673 ], "response_mean_nll": 0.09195867602596103, "schema": "uor-r4.stack-dialogue-development/1", "scope": "Token means over the selected development responses, not the full corpus or an equal-source mean.", "selected_responses": 32, "supervised_targets": 327 }, "flock_and_pointer": { "final_pointer_diagnostics": { "definitions": "over the scored targets: mean_gate is the mean gate g_t; pointer_hit_rate is the fraction whose most attended source (lowest position on a tie) holds the target token; target_reachable_rate is the fraction whose target token is held by any source with attention (the hit rate's ceiling); mean_copy_mass is the mean p_copy(target) before the gate; gate_bce is the mean of -log g_t on reachable targets and -log(1 - g_t) on the others, and pointer_nll the sum of -log p_copy(target) over reachable targets divided by all scored targets (the two copy-gate supervision terms, reported whether or not a run trains them)", "gate_bce": 5.255796121810704, "mean_copy_mass": 0.21853918230027355, "mean_gate": 0.011816133449776159, "pointer_hit_rate": 0.26605504587155965, "pointer_nll": 1.2457404604805489, "scored_targets": 327, "target_reachable_rate": 0.5474006116207951 }, "pointer": { "dim": 32, "init_seed": 1 }, "pointer_gate_supervision": null, "pointer_head_added_to_init": true, "pointer_init_seed": 1, "pointer_init_seed_source": "the saved model's seed", "pointer_parameters": 99841, "pointer_route": null, "pointer_score": "dot", "pointer_select": null, "read_binding_supervision": null, "scope": "Flock selection of the reads (sink at position 0, last WINDOW positions and the K best-scoring other sources per read row, by the shared crate::flock selector; unkept sources weigh and receive exactly 0; the NoRead slot stays outside the selection) applies to every read of the model and never to the pointer. The pointer head scores each source of the window with its own score (`pointer_score`: dot, or the fused read's Lorentz form with a learned scale) and softmaxes over the sources its own selection keeps (`pointer_select`: none keeps all, top:1 is the single-source pointer), copies the input tokens at the attended positions, and a gate g = sigmoid(w.h + b) (b starts at -2) mixes that with the ordinary distribution; when no kept source holds a target the mixture is (1 - g) softmax alone, with no floor. `response_mean_nll`, the losses and the greedy replies are the mixture's. `pointer` diagnostics are over the scored dev targets. Neither a flock nor a pointer selection or route has a D11 port; `export` writes a pointer that keeps every source (both integer engines serve its mixture), and `qat=true` refuses a pointer. Offline float training only; not a served or quality result.", "select": null }, "initial_development": { "conditioning": "full_original_prefix", "eos_targets": 32, "first_four_response_targets_mean_nll": 2.7206015877480016, "first_four_targets": 128, "per_source": [ { "eligible_responses": 1223, "eos_targets": 32, "first_four_response_targets_mean_nll": 2.7206015877480016, "first_four_targets": 128, "label": "m-world-v2.train", "pointer": { "definitions": "over the scored targets: mean_gate is the mean gate g_t; pointer_hit_rate is the fraction whose most attended source (lowest position on a tie) holds the target token; target_reachable_rate is the fraction whose target token is held by any source with attention (the hit rate's ceiling); mean_copy_mass is the mean p_copy(target) before the gate; gate_bce is the mean of -log g_t on reachable targets and -log(1 - g_t) on the others, and pointer_nll the sum of -log p_copy(target) over reachable targets divided by all scored targets (the two copy-gate supervision terms, reported whether or not a run trains them)", "gate_bce": 1.221729234284567, "mean_copy_mass": 0.03141511729955104, "mean_gate": 0.1192029220221181, "pointer_hit_rate": 0.024464831804281346, "pointer_nll": 2.95247560357175, "scored_targets": 327, "target_reachable_rate": 0.5474006116207951 }, "response_ids": [ 962, 1138, 101, 1102, 613, 671, 283, 1185, 421, 484, 1129, 639, 170, 1063, 677, 200, 143, 714, 424, 928, 880, 1015, 740, 999, 427, 420, 395, 76, 1221, 936, 1104, 673 ], "response_mean_nll": 2.2791918762378933, "selected_responses": 32, "source_index": 0, "supervised_targets": 327 } ], "pointer": { "definitions": "over the scored targets: mean_gate is the mean gate g_t; pointer_hit_rate is the fraction whose most attended source (lowest position on a tie) holds the target token; target_reachable_rate is the fraction whose target token is held by any source with attention (the hit rate's ceiling); mean_copy_mass is the mean p_copy(target) before the gate; gate_bce is the mean of -log g_t on reachable targets and -log(1 - g_t) on the others, and pointer_nll the sum of -log p_copy(target) over reachable targets divided by all scored targets (the two copy-gate supervision terms, reported whether or not a run trains them)", "gate_bce": 1.221729234284567, "mean_copy_mass": 0.03141511729955104, "mean_gate": 0.1192029220221181, "pointer_hit_rate": 0.024464831804281346, "pointer_nll": 2.95247560357175, "scored_targets": 327, "target_reachable_rate": 0.5474006116207951 }, "response_ids": [ 962, 1138, 101, 1102, 613, 671, 283, 1185, 421, 484, 1129, 639, 170, 1063, 677, 200, 143, 714, 424, 928, 880, 1015, 740, 999, 427, 420, 395, 76, 1221, 936, 1104, 673 ], "response_mean_nll": 2.2791918762378933, "schema": "uor-r4.stack-dialogue-development/1", "scope": "Token means over the selected development responses, not the full corpus or an equal-source mean.", "selected_responses": 32, "supervised_targets": 327 }, "inputs": { "dev": [ { "bytes": 85776, "path": "/b50data/ft/dev/tokens.u16", "sha256": "0919e90f953e988edf05cbb8a69e24bafda0cb8b8ddbbb5d1e9d808c81a6555b" }, { "bytes": 42856, "path": "/b50data/ft/dev/response_mask.u8", "sha256": "a838bb1a67e96572cb10a86b5a3587dda25f0190eaf85133026e04c281afb078" }, { "bytes": 5398, "path": "/b50data/ft/dev/manifest.json", "sha256": "74a9158604336ef8079e13e615508ea4cd56b2253ce020b3335979cb152f847c" } ], "init": { "bytes": 855748328, "path": "/uor-r4/planA/model/model.safetensors", "sha256": "2b425ccdb2b5eae4d3f8dab924ee09cb6d167cb7aa65188de7cf9d24ef6d6586" }, "requests": null, "tokenizer": { "bytes": 109457, "path": "/b50data/tokenizer.json", "sha256": "d36d3e8700a123e620012df77de195f244fbdb4d05d9e1aa7e77fa9407590f89" }, "train": [ { "bytes": 346497520, "path": "/b50data/ft/mixB/tokens.u16", "sha256": "2f7202bcfa6c804863dade2e6e27a65e8093018868c40807913447e5fc00671c" }, { "bytes": 173248728, "path": "/b50data/ft/mixB/response_mask.u8", "sha256": "9baabe5c0e1590613bdfc19c507cfde9c2069bdf7727ed006698ca83806727e8" }, { "bytes": 2538, "path": "/b50data/ft/mixB/manifest.json", "sha256": "6913797cd8dd8d0e452f32885cd6c437e5ab0a33d2f52ddfccfb595ba4d97a3a" } ] }, "model_sha256": "11b84ba7d978a52c3808a519e859e74182b3ebc94713581530423d85db120d4b", "parameters": 214031977, "protocol": { "bos_id": 0, "eos_id": 1, "schema": "uor-r4.literal-role-dialogue/2", "tokenizer_cid": "blake3:3f42bcfce7728512076549c63b88387e13c8156fe35c0f91d9b112439f3739cc", "unk_id": 2 }, "replies": null, "resumed_from": null, "sampler": "uor-r4.dialogue-response-uniform/splitmix64-counter-rejection-v1", "schema": "uor-r4.geometric-stack-dialogue-run/1", "scope": "Offline response learning of the geometric stack on the prepared dialogue corpus; development NLL on the fixed source-stratified panel under full original prefixes; replies from the float model. No integer serving or quality qualification.", "settings": { "batch": 16, "checkpoint_every": 8000, "clip": 1.0, "data_seed": 1, "dev": [ "/b50data/ft/dev/tokens.u16", "/b50data/ft/dev/response_mask.u8", "/b50data/ft/dev/manifest.json" ], "dev_per_source": 32, "dev_seed": 20260930, "eval_every": 2000, "init": "/uor-r4/planA/model", "key_shift": "false", "lr": 0.00015, "max_new_tokens": 32, "min_lr": 0.1, "pointer": "32", "policy": "full_prefix", "precision": "f32", "protocol": 2, "qat": false, "qat_codec": null, "read_lineage": null, "requests": null, "steps": 8000, "tokenizer": "/b50data/tokenizer.json", "train": [ "/b50data/ft/mixB/tokens.u16", "/b50data/ft/mixB/response_mask.u8", "/b50data/ft/mixB/manifest.json" ], "transport_snap": null, "warmup": 200, "weight_decay": 0.1 }, "step_seconds": { "first": 0.182019601, "max_after_first": 0.194504857, "mean_after_first": 0.16353614577672249, "min_after_first": 0.076475261, "steps": 8000 }, "stopped_early": false, "supervised_target_visits": 5155535, "tensor_positions_this_process": 43730336, "tf32": true, "threads": "8", "train_population": { "corpus_tokens": 173248728, "documents": 521377, "eligible_response_tokens": 34985344, "eligible_responses": 876603, "excluded_over_context": 97667, "identical_prefixes": 0, "response_runs": 974270, "response_tokens": 88184122, "sources": [ { "documents": 208625, "eligible_response_tokens": 8926398, "eligible_responses": 511577, "excluded_over_context": 97667, "identical_prefixes": 0, "label": "balp", "response_runs": 609244, "response_tokens": 62125176, "source_index": 0 }, { "documents": 113768, "eligible_response_tokens": 1756266, "eligible_responses": 164767, "excluded_over_context": 0, "identical_prefixes": 0, "label": "dialogue-recall-v2", "response_runs": 164767, "response_tokens": 1756266, "source_index": 1 }, { "documents": 198984, "eligible_response_tokens": 24302680, "eligible_responses": 200259, "excluded_over_context": 0, "identical_prefixes": 0, "label": "smoltalk", "response_runs": 200259, "response_tokens": 24302680, "source_index": 2 } ] }, "train_seconds": 1308.3076496690032 }