@f-inverse/jammi-client 0.25.0 → 0.26.0

package/dist/gen/jammi/v1/training_pb.d.ts

@@ -0,0 +1,1217 @@
+import type { GenEnum, GenFile, GenMessage, GenService } from "@bufbuild/protobuf/codegenv2";
+import type { ModelTask } from "./inference_pb";
+import type { Message } from "@bufbuild/protobuf";
+/**
+ * Describes the file jammi/v1/training.proto.
+ */
+export declare const file_jammi_v1_training: GenFile;
+/**
+ * Embedding-loss function. Mirrors the engine's `EmbeddingLoss`; the two
+ * margin/temperature-carrying variants carry their scalar parameter.
+ *
+ * @generated from message jammi.v1.training.EmbeddingLoss
+ */
+export type EmbeddingLoss = Message<"jammi.v1.training.EmbeddingLoss"> & {
+    /**
+     * @generated from oneof jammi.v1.training.EmbeddingLoss.loss
+     */
+    loss: {
+        /**
+         * CoSENT: cross-entropy over cosine-similarity ordering. No parameter.
+         *
+         * @generated from field: jammi.v1.training.EmbeddingLoss.CoSent co_sent = 1;
+         */
+        value: EmbeddingLoss_CoSent;
+        case: "coSent";
+    } | {
+        /**
+         * Triplet loss with the given margin.
+         *
+         * @generated from field: jammi.v1.training.EmbeddingLoss.Triplet triplet = 2;
+         */
+        value: EmbeddingLoss_Triplet;
+        case: "triplet";
+    } | {
+        /**
+         * InfoNCE with in-batch negatives at the given temperature.
+         *
+         * @generated from field: jammi.v1.training.EmbeddingLoss.MultipleNegativesRanking multiple_negatives_ranking = 3;
+         */
+        value: EmbeddingLoss_MultipleNegativesRanking;
+        case: "multipleNegativesRanking";
+    } | {
+        /**
+         * AnglE: angle-difference objective in complex space. No parameter.
+         *
+         * @generated from field: jammi.v1.training.EmbeddingLoss.AnglE angle = 4;
+         */
+        value: EmbeddingLoss_AnglE;
+        case: "angle";
+    } | {
+        /**
+         * cosine-MSE: regress scaled cosine onto a graded score. No parameter.
+         *
+         * @generated from field: jammi.v1.training.EmbeddingLoss.CosineMse cosine_mse = 5;
+         */
+        value: EmbeddingLoss_CosineMse;
+        case: "cosineMse";
+    } | {
+        case: undefined;
+        value?: undefined;
+    };
+};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.
+ * Use `create(EmbeddingLossSchema)` to create a new message.
+ */
+export declare const EmbeddingLossSchema: GenMessage<EmbeddingLoss>;
+/**
+ * @generated from message jammi.v1.training.EmbeddingLoss.CoSent
+ */
+export type EmbeddingLoss_CoSent = Message<"jammi.v1.training.EmbeddingLoss.CoSent"> & {};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.CoSent.
+ * Use `create(EmbeddingLoss_CoSentSchema)` to create a new message.
+ */
+export declare const EmbeddingLoss_CoSentSchema: GenMessage<EmbeddingLoss_CoSent>;
+/**
+ * @generated from message jammi.v1.training.EmbeddingLoss.Triplet
+ */
+export type EmbeddingLoss_Triplet = Message<"jammi.v1.training.EmbeddingLoss.Triplet"> & {
+    /**
+     * @generated from field: double margin = 1;
+     */
+    margin: number;
+};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.Triplet.
+ * Use `create(EmbeddingLoss_TripletSchema)` to create a new message.
+ */
+export declare const EmbeddingLoss_TripletSchema: GenMessage<EmbeddingLoss_Triplet>;
+/**
+ * @generated from message jammi.v1.training.EmbeddingLoss.MultipleNegativesRanking
+ */
+export type EmbeddingLoss_MultipleNegativesRanking = Message<"jammi.v1.training.EmbeddingLoss.MultipleNegativesRanking"> & {
+    /**
+     * @generated from field: double temperature = 1;
+     */
+    temperature: number;
+};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.MultipleNegativesRanking.
+ * Use `create(EmbeddingLoss_MultipleNegativesRankingSchema)` to create a new message.
+ */
+export declare const EmbeddingLoss_MultipleNegativesRankingSchema: GenMessage<EmbeddingLoss_MultipleNegativesRanking>;
+/**
+ * @generated from message jammi.v1.training.EmbeddingLoss.AnglE
+ */
+export type EmbeddingLoss_AnglE = Message<"jammi.v1.training.EmbeddingLoss.AnglE"> & {};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.AnglE.
+ * Use `create(EmbeddingLoss_AnglESchema)` to create a new message.
+ */
+export declare const EmbeddingLoss_AnglESchema: GenMessage<EmbeddingLoss_AnglE>;
+/**
+ * @generated from message jammi.v1.training.EmbeddingLoss.CosineMse
+ */
+export type EmbeddingLoss_CosineMse = Message<"jammi.v1.training.EmbeddingLoss.CosineMse"> & {};
+/**
+ * Describes the message jammi.v1.training.EmbeddingLoss.CosineMse.
+ * Use `create(EmbeddingLoss_CosineMseSchema)` to create a new message.
+ */
+export declare const EmbeddingLoss_CosineMseSchema: GenMessage<EmbeddingLoss_CosineMse>;
+/**
+ * Proper-scoring objective for a distributional regression head (S18). Mirrors
+ * the engine's `RegressionLoss`; the β-NLL variant carries its weighting
+ * exponent. Pinball trains the quantile head over `quantile_levels`; the others
+ * train the parametric Gaussian head.
+ *
+ * @generated from message jammi.v1.training.RegressionLoss
+ */
+export type RegressionLoss = Message<"jammi.v1.training.RegressionLoss"> & {
+    /**
+     * @generated from oneof jammi.v1.training.RegressionLoss.loss
+     */
+    loss: {
+        /**
+         * Gaussian NLL — the plain heteroscedastic mean-variance objective.
+         *
+         * @generated from field: jammi.v1.training.RegressionLoss.GaussianNll gaussian_nll = 1;
+         */
+        value: RegressionLoss_GaussianNll;
+        case: "gaussianNll";
+    } | {
+        /**
+         * β-NLL with the given variance-weighting exponent (Seitzer 2022).
+         *
+         * @generated from field: jammi.v1.training.RegressionLoss.BetaNll beta_nll = 2;
+         */
+        value: RegressionLoss_BetaNll;
+        case: "betaNll";
+    } | {
+        /**
+         * Closed-form Gaussian CRPS. No parameter.
+         *
+         * @generated from field: jammi.v1.training.RegressionLoss.Crps crps = 3;
+         */
+        value: RegressionLoss_Crps;
+        case: "crps";
+    } | {
+        /**
+         * Pinball / quantile loss over the configured levels. No parameter.
+         *
+         * @generated from field: jammi.v1.training.RegressionLoss.Pinball pinball = 4;
+         */
+        value: RegressionLoss_Pinball;
+        case: "pinball";
+    } | {
+        case: undefined;
+        value?: undefined;
+    };
+};
+/**
+ * Describes the message jammi.v1.training.RegressionLoss.
+ * Use `create(RegressionLossSchema)` to create a new message.
+ */
+export declare const RegressionLossSchema: GenMessage<RegressionLoss>;
+/**
+ * @generated from message jammi.v1.training.RegressionLoss.GaussianNll
+ */
+export type RegressionLoss_GaussianNll = Message<"jammi.v1.training.RegressionLoss.GaussianNll"> & {};
+/**
+ * Describes the message jammi.v1.training.RegressionLoss.GaussianNll.
+ * Use `create(RegressionLoss_GaussianNllSchema)` to create a new message.
+ */
+export declare const RegressionLoss_GaussianNllSchema: GenMessage<RegressionLoss_GaussianNll>;
+/**
+ * @generated from message jammi.v1.training.RegressionLoss.BetaNll
+ */
+export type RegressionLoss_BetaNll = Message<"jammi.v1.training.RegressionLoss.BetaNll"> & {
+    /**
+     * @generated from field: double beta = 1;
+     */
+    beta: number;
+};
+/**
+ * Describes the message jammi.v1.training.RegressionLoss.BetaNll.
+ * Use `create(RegressionLoss_BetaNllSchema)` to create a new message.
+ */
+export declare const RegressionLoss_BetaNllSchema: GenMessage<RegressionLoss_BetaNll>;
+/**
+ * @generated from message jammi.v1.training.RegressionLoss.Crps
+ */
+export type RegressionLoss_Crps = Message<"jammi.v1.training.RegressionLoss.Crps"> & {};
+/**
+ * Describes the message jammi.v1.training.RegressionLoss.Crps.
+ * Use `create(RegressionLoss_CrpsSchema)` to create a new message.
+ */
+export declare const RegressionLoss_CrpsSchema: GenMessage<RegressionLoss_Crps>;
+/**
+ * @generated from message jammi.v1.training.RegressionLoss.Pinball
+ */
+export type RegressionLoss_Pinball = Message<"jammi.v1.training.RegressionLoss.Pinball"> & {};
+/**
+ * Describes the message jammi.v1.training.RegressionLoss.Pinball.
+ * Use `create(RegressionLoss_PinballSchema)` to create a new message.
+ */
+export declare const RegressionLoss_PinballSchema: GenMessage<RegressionLoss_Pinball>;
+/**
+ * 0-based encoder layer indices that receive LoRA adapters. Wrapper message so
+ * "no restriction" (field absent → all layers) is distinguishable from "an
+ * empty restriction"; mirrors the engine's `Option<Vec<usize>>`.
+ *
+ * @generated from message jammi.v1.training.LayersToTransform
+ */
+export type LayersToTransform = Message<"jammi.v1.training.LayersToTransform"> & {
+    /**
+     * @generated from field: repeated uint32 layers = 1;
+     */
+    layers: number[];
+};
+/**
+ * Describes the message jammi.v1.training.LayersToTransform.
+ * Use `create(LayersToTransformSchema)` to create a new message.
+ */
+export declare const LayersToTransformSchema: GenMessage<LayersToTransform>;
+/**
+ * Hard-negative mining configuration. Mirrors the engine's
+ * `HardNegativeConfig`. When `mine` is set the trainer mines hard negatives
+ * from its own ANN index, excluding the positive's `exclude_hops`-hop
+ * neighbourhood and re-mining every `refresh_every` epochs.
+ *
+ * @generated from message jammi.v1.training.HardNegativeConfig
+ */
+export type HardNegativeConfig = Message<"jammi.v1.training.HardNegativeConfig"> & {
+    /**
+     * @generated from field: bool mine = 1;
+     */
+    mine: boolean;
+    /**
+     * @generated from field: uint32 k = 2;
+     */
+    k: number;
+    /**
+     * @generated from field: uint32 exclude_hops = 3;
+     */
+    excludeHops: number;
+    /**
+     * @generated from field: uint32 refresh_every = 4;
+     */
+    refreshEvery: number;
+};
+/**
+ * Describes the message jammi.v1.training.HardNegativeConfig.
+ * Use `create(HardNegativeConfigSchema)` to create a new message.
+ */
+export declare const HardNegativeConfigSchema: GenMessage<HardNegativeConfig>;
+/**
+ * Configuration for a fine-tuning job. Mirrors the engine's `FineTuneConfig`
+ * field for field; every field is read by the trainer or the LoRA builder.
+ * Every scalar knob is `optional` so an unset field is distinguishable from a
+ * legal `0`/`false`: the server reads an unset field as "apply the engine
+ * default" (`FineTuneConfig::default()`) and a present field as an override,
+ * making the engine the single source of default values. `embedding_loss`/
+ * `classification_loss` are unset to let the engine auto-select from the data
+ * format; `layers_to_transform` is unset for "all layers". The enum-typed
+ * fields default to the engine's default variant when left UNSPECIFIED.
+ *
+ * @generated from message jammi.v1.training.FineTuneConfig
+ */
+export type FineTuneConfig = Message<"jammi.v1.training.FineTuneConfig"> & {
+    /**
+     * @generated from field: optional uint32 lora_rank = 1;
+     */
+    loraRank?: number | undefined;
+    /**
+     * @generated from field: optional double lora_alpha = 2;
+     */
+    loraAlpha?: number | undefined;
+    /**
+     * @generated from field: optional double lora_dropout = 3;
+     */
+    loraDropout?: number | undefined;
+    /**
+     * @generated from field: optional double learning_rate = 4;
+     */
+    learningRate?: number | undefined;
+    /**
+     * @generated from field: optional uint32 epochs = 5;
+     */
+    epochs?: number | undefined;
+    /**
+     * @generated from field: optional uint32 batch_size = 6;
+     */
+    batchSize?: number | undefined;
+    /**
+     * @generated from field: optional uint32 max_seq_length = 7;
+     */
+    maxSeqLength?: number | undefined;
+    /**
+     * Unset = auto-select from the data format.
+     *
+     * @generated from field: optional jammi.v1.training.EmbeddingLoss embedding_loss = 8;
+     */
+    embeddingLoss?: EmbeddingLoss | undefined;
+    /**
+     * Unset = auto-select from the data format.
+     *
+     * @generated from field: optional jammi.v1.training.ClassificationLoss classification_loss = 9;
+     */
+    classificationLoss?: ClassificationLoss | undefined;
+    /**
+     * @generated from field: optional uint32 gradient_accumulation_steps = 10;
+     */
+    gradientAccumulationSteps?: number | undefined;
+    /**
+     * @generated from field: optional double validation_fraction = 11;
+     */
+    validationFraction?: number | undefined;
+    /**
+     * @generated from field: optional uint32 early_stopping_patience = 12;
+     */
+    earlyStoppingPatience?: number | undefined;
+    /**
+     * @generated from field: optional uint32 warmup_steps = 13;
+     */
+    warmupSteps?: number | undefined;
+    /**
+     * @generated from field: jammi.v1.training.LrSchedule lr_schedule = 14;
+     */
+    lrSchedule: LrSchedule;
+    /**
+     * @generated from field: jammi.v1.training.EarlyStoppingMetric early_stopping_metric = 15;
+     */
+    earlyStoppingMetric: EarlyStoppingMetric;
+    /**
+     * Layer-name suffixes that receive LoRA adapters. Empty = projection head
+     * over the frozen base model; non-empty = LoRA inside the encoder.
+     *
+     * @generated from field: repeated string target_modules = 16;
+     */
+    targetModules: string[];
+    /**
+     * Unset = all layers.
+     *
+     * @generated from field: optional jammi.v1.training.LayersToTransform layers_to_transform = 17;
+     */
+    layersToTransform?: LayersToTransform | undefined;
+    /**
+     * @generated from field: optional bool use_rslora = 18;
+     */
+    useRslora?: boolean | undefined;
+    /**
+     * Per-module rank overrides keyed by module-name substring.
+     *
+     * @generated from field: map<string, uint32> rank_pattern = 19;
+     */
+    rankPattern: {
+        [key: string]: number;
+    };
+    /**
+     * @generated from field: jammi.v1.training.LoraInitMode init_lora_weights = 20;
+     */
+    initLoraWeights: LoraInitMode;
+    /**
+     * @generated from field: jammi.v1.training.BackboneDtype backbone_dtype = 21;
+     */
+    backboneDtype: BackboneDtype;
+    /**
+     * @generated from field: optional double weight_decay = 22;
+     */
+    weightDecay?: number | undefined;
+    /**
+     * @generated from field: optional double max_grad_norm = 23;
+     */
+    maxGradNorm?: number | undefined;
+    /**
+     * GradCache: two-pass in-batch-negative training that enlarges the effective
+     * negative pool without the memory cost. Only affects MultipleNegativesRanking.
+     *
+     * @generated from field: optional bool cached = 24;
+     */
+    cached?: boolean | undefined;
+    /**
+     * Hard-negative mining over jammi's own ANN index. Field absent = mining off.
+     *
+     * @generated from field: optional jammi.v1.training.HardNegativeConfig hard_negatives = 25;
+     */
+    hardNegatives?: HardNegativeConfig | undefined;
+    /**
+     * Matryoshka prefix dimensions. Empty = full dimension only.
+     *
+     * @generated from field: repeated uint32 matryoshka_dims = 26;
+     */
+    matryoshkaDims: number[];
+    /**
+     * Unset = the engine default (β-NLL, β=0.5). The Pinball variant trains the
+     * quantile head over `quantile_levels`.
+     *
+     * @generated from field: optional jammi.v1.training.RegressionLoss regression_loss = 27;
+     */
+    regressionLoss?: RegressionLoss | undefined;
+    /**
+     * Quantile levels for a pinball-trained regression head, ascending in (0, 1).
+     * Empty for the parametric Gaussian objectives.
+     *
+     * @generated from field: repeated double quantile_levels = 28;
+     */
+    quantileLevels: number[];
+};
+/**
+ * Describes the message jammi.v1.training.FineTuneConfig.
+ * Use `create(FineTuneConfigSchema)` to create a new message.
+ */
+export declare const FineTuneConfigSchema: GenMessage<FineTuneConfig>;
+/**
+ * The two sources and column bindings a graph fine-tune reads. Mirrors the
+ * engine's `GraphFineTuneSources` field for field.
+ *
+ * @generated from message jammi.v1.training.GraphFineTuneSources
+ */
+export type GraphFineTuneSources = Message<"jammi.v1.training.GraphFineTuneSources"> & {
+    /**
+     * Catalog source holding the node text.
+     *
+     * @generated from field: string node_source = 1;
+     */
+    nodeSource: string;
+    /**
+     * Column in `node_source` holding the node id (edge endpoints join to it).
+     *
+     * @generated from field: string id_column = 2;
+     */
+    idColumn: string;
+    /**
+     * Column in `node_source` holding the text the encoder embeds.
+     *
+     * @generated from field: string text_column = 3;
+     */
+    textColumn: string;
+    /**
+     * Catalog source holding the edges.
+     *
+     * @generated from field: string edge_source = 4;
+     */
+    edgeSource: string;
+    /**
+     * Column in `edge_source` holding the edge source endpoint.
+     *
+     * @generated from field: string src_column = 5;
+     */
+    srcColumn: string;
+    /**
+     * Column in `edge_source` holding the edge destination endpoint.
+     *
+     * @generated from field: string dst_column = 6;
+     */
+    dstColumn: string;
+    /**
+     * Provenance every edge in `edge_source` carries.
+     *
+     * @generated from field: jammi.v1.training.EdgeProvenance provenance = 7;
+     */
+    provenance: EdgeProvenance;
+};
+/**
+ * Describes the message jammi.v1.training.GraphFineTuneSources.
+ * Use `create(GraphFineTuneSourcesSchema)` to create a new message.
+ */
+export declare const GraphFineTuneSourcesSchema: GenMessage<GraphFineTuneSources>;
+/**
+ * node2vec walk + structure-aware negative-sampling knobs. Mirrors the engine's
+ * `GraphSampleConfig` field for field, including the deterministic `seed`.
+ *
+ * @generated from message jammi.v1.training.GraphSampleConfig
+ */
+export type GraphSampleConfig = Message<"jammi.v1.training.GraphSampleConfig"> & {
+    /**
+     * @generated from field: uint32 walk_length = 1;
+     */
+    walkLength: number;
+    /**
+     * @generated from field: uint32 walks_per_node = 2;
+     */
+    walksPerNode: number;
+    /**
+     * @generated from field: double return_p = 3;
+     */
+    returnP: number;
+    /**
+     * @generated from field: double in_out_q = 4;
+     */
+    inOutQ: number;
+    /**
+     * @generated from field: uint32 hard_negatives = 5;
+     */
+    hardNegatives: number;
+    /**
+     * @generated from field: uint32 exclude_hops = 6;
+     */
+    excludeHops: number;
+    /**
+     * @generated from field: uint32 min_negatives = 7;
+     */
+    minNegatives: number;
+    /**
+     * @generated from field: uint64 seed = 8;
+     */
+    seed: bigint;
+};
+/**
+ * Describes the message jammi.v1.training.GraphSampleConfig.
+ * Use `create(GraphSampleConfigSchema)` to create a new message.
+ */
+export declare const GraphSampleConfigSchema: GenMessage<GraphSampleConfig>;
+/**
+ * The predictive-distribution head + its proper-scoring objective. Mirrors the
+ * engine's `PredictiveHead`: a Gaussian head trained against a Gaussian
+ * objective, or a quantile head trained with pinball over its levels.
+ *
+ * @generated from message jammi.v1.training.PredictiveHead
+ */
+export type PredictiveHead = Message<"jammi.v1.training.PredictiveHead"> & {
+    /**
+     * @generated from oneof jammi.v1.training.PredictiveHead.head
+     */
+    head: {
+        /**
+         * @generated from field: jammi.v1.training.PredictiveHead.Gaussian gaussian = 1;
+         */
+        value: PredictiveHead_Gaussian;
+        case: "gaussian";
+    } | {
+        /**
+         * @generated from field: jammi.v1.training.PredictiveHead.Quantile quantile = 2;
+         */
+        value: PredictiveHead_Quantile;
+        case: "quantile";
+    } | {
+        case: undefined;
+        value?: undefined;
+    };
+};
+/**
+ * Describes the message jammi.v1.training.PredictiveHead.
+ * Use `create(PredictiveHeadSchema)` to create a new message.
+ */
+export declare const PredictiveHeadSchema: GenMessage<PredictiveHead>;
+/**
+ * A `(mean, raw_std)` Gaussian head with its proper-scoring objective.
+ *
+ * @generated from message jammi.v1.training.PredictiveHead.Gaussian
+ */
+export type PredictiveHead_Gaussian = Message<"jammi.v1.training.PredictiveHead.Gaussian"> & {
+    /**
+     * @generated from field: jammi.v1.training.GaussianObjective objective = 1;
+     */
+    objective?: GaussianObjective | undefined;
+};
+/**
+ * Describes the message jammi.v1.training.PredictiveHead.Gaussian.
+ * Use `create(PredictiveHead_GaussianSchema)` to create a new message.
+ */
+export declare const PredictiveHead_GaussianSchema: GenMessage<PredictiveHead_Gaussian>;
+/**
+ * A `levels`-wide quantile head trained with the pinball objective.
+ *
+ * @generated from message jammi.v1.training.PredictiveHead.Quantile
+ */
+export type PredictiveHead_Quantile = Message<"jammi.v1.training.PredictiveHead.Quantile"> & {
+    /**
+     * Ascending quantile levels in (0, 1); the head width is their count.
+     *
+     * @generated from field: repeated double levels = 1;
+     */
+    levels: number[];
+};
+/**
+ * Describes the message jammi.v1.training.PredictiveHead.Quantile.
+ * Use `create(PredictiveHead_QuantileSchema)` to create a new message.
+ */
+export declare const PredictiveHead_QuantileSchema: GenMessage<PredictiveHead_Quantile>;
+/**
+ * Which proper score a Gaussian predictive head trains against. Mirrors the
+ * engine's `GaussianObjective`: (β-)NLL carrying its β, or closed-form CRPS.
+ *
+ * @generated from message jammi.v1.training.GaussianObjective
+ */
+export type GaussianObjective = Message<"jammi.v1.training.GaussianObjective"> & {
+    /**
+     * @generated from oneof jammi.v1.training.GaussianObjective.objective
+     */
+    objective: {
+        /**
+         * (β-)Gaussian NLL; `beta` in [0, 1], 0 the plain heteroscedastic NLL.
+         *
+         * @generated from field: jammi.v1.training.GaussianObjective.Nll nll = 1;
+         */
+        value: GaussianObjective_Nll;
+        case: "nll";
+    } | {
+        /**
+         * Closed-form Gaussian CRPS — collapse-resistant, outcome-unit.
+         *
+         * @generated from field: jammi.v1.training.GaussianObjective.Crps crps = 2;
+         */
+        value: GaussianObjective_Crps;
+        case: "crps";
+    } | {
+        case: undefined;
+        value?: undefined;
+    };
+};
+/**
+ * Describes the message jammi.v1.training.GaussianObjective.
+ * Use `create(GaussianObjectiveSchema)` to create a new message.
+ */
+export declare const GaussianObjectiveSchema: GenMessage<GaussianObjective>;
+/**
+ * @generated from message jammi.v1.training.GaussianObjective.Nll
+ */
+export type GaussianObjective_Nll = Message<"jammi.v1.training.GaussianObjective.Nll"> & {
+    /**
+     * @generated from field: double beta = 1;
+     */
+    beta: number;
+};
+/**
+ * Describes the message jammi.v1.training.GaussianObjective.Nll.
+ * Use `create(GaussianObjective_NllSchema)` to create a new message.
+ */
+export declare const GaussianObjective_NllSchema: GenMessage<GaussianObjective_Nll>;
+/**
+ * @generated from message jammi.v1.training.GaussianObjective.Crps
+ */
+export type GaussianObjective_Crps = Message<"jammi.v1.training.GaussianObjective.Crps"> & {};
+/**
+ * Describes the message jammi.v1.training.GaussianObjective.Crps.
+ * Use `create(GaussianObjective_CrpsSchema)` to create a new message.
+ */
+export declare const GaussianObjective_CrpsSchema: GenMessage<GaussianObjective_Crps>;
+/**
+ * The episodic in-context-predictor meta-training specification. Mirrors the
+ * engine's `ContextPredictorTrainConfig` field for field.
+ *
+ * @generated from message jammi.v1.training.ContextPredictorTrainConfig
+ */
+export type ContextPredictorTrainConfig = Message<"jammi.v1.training.ContextPredictorTrainConfig"> & {
+    /**
+     * The model id the trained predictor registers under in the catalog.
+     *
+     * @generated from field: string model_id = 1;
+     */
+    modelId: string;
+    /**
+     * Which curated architecture to train.
+     *
+     * @generated from field: jammi.v1.training.ContextArchitecture architecture = 2;
+     */
+    architecture: ContextArchitecture;
+    /**
+     * The source's key column — the per-row identity episodes sample by.
+     *
+     * @generated from field: string key_column = 3;
+     */
+    keyColumn: string;
+    /**
+     * The column whose distinct values are the tasks (the episodic partition).
+     *
+     * @generated from field: string task_column = 4;
+     */
+    taskColumn: string;
+    /**
+     * The column carrying the scalar outcome `y` the predictor regresses.
+     *
+     * @generated from field: string value_column = 5;
+     */
+    valueColumn: string;
+    /**
+     * Context size `k`: the retrieval over-fetches and pads the episode to this.
+     *
+     * @generated from field: uint32 context_k = 6;
+     */
+    contextK: number;
+    /**
+     * Hidden width of the predictor's MLPs / transformer model dimension.
+     *
+     * @generated from field: uint32 hidden_dim = 7;
+     */
+    hiddenDim: number;
+    /**
+     * Attention heads (AttnCnp, Tnp). Must divide `hidden_dim`.
+     *
+     * @generated from field: uint32 num_heads = 8;
+     */
+    numHeads: number;
+    /**
+     * Transformer layers (Tnp).
+     *
+     * @generated from field: uint32 num_layers = 9;
+     */
+    numLayers: number;
+    /**
+     * The predictive-distribution head + its proper-scoring objective.
+     *
+     * @generated from field: jammi.v1.training.PredictiveHead head = 10;
+     */
+    head?: PredictiveHead | undefined;
+    /**
+     * Passes over the sampled train episodes.
+     *
+     * @generated from field: uint32 epochs = 11;
+     */
+    epochs: number;
+    /**
+     * AdamW learning rate.
+     *
+     * @generated from field: double learning_rate = 12;
+     */
+    learningRate: number;
+    /**
+     * Global-L2 gradient-clip norm; <= 0.0 disables clipping.
+     *
+     * @generated from field: double grad_clip = 13;
+     */
+    gradClip: number;
+    /**
+     * Fraction of the distinct tasks held out for the test split, in (0, 1).
+     *
+     * @generated from field: double test_task_fraction = 14;
+     */
+    testTaskFraction: number;
+    /**
+     * Minimum distinct task count the meta-dataset must carry.
+     *
+     * @generated from field: uint32 min_task_count = 15;
+     */
+    minTaskCount: number;
+    /**
+     * Seed for the deterministic train/test task partition.
+     *
+     * @generated from field: uint64 seed = 16;
+     */
+    seed: bigint;
+};
+/**
+ * Describes the message jammi.v1.training.ContextPredictorTrainConfig.
+ * Use `create(ContextPredictorTrainConfigSchema)` to create a new message.
+ */
+export declare const ContextPredictorTrainConfigSchema: GenMessage<ContextPredictorTrainConfig>;
+/**
+ * A column-source contrastive / classification / regression fine-tune. Mirrors
+ * the engine's `TrainingSpec::FineTune` variant.
+ *
+ * @generated from message jammi.v1.training.FineTuneSpec
+ */
+export type FineTuneSpec = Message<"jammi.v1.training.FineTuneSpec"> & {
+    /**
+     * Registered source whose rows the model is fine-tuned on.
+     *
+     * @generated from field: string source = 1;
+     */
+    source: string;
+    /**
+     * Columns selected into the training loader (the format is detected from
+     * these names).
+     *
+     * @generated from field: repeated string columns = 2;
+     */
+    columns: string[];
+    /**
+     * Adapter method.
+     *
+     * @generated from field: jammi.v1.training.FineTuneMethod method = 3;
+     */
+    method: FineTuneMethod;
+    /**
+     * Task the fine-tuned model performs.
+     *
+     * @generated from field: jammi.v1.inference.ModelTask task = 4;
+     */
+    task: ModelTask;
+};
+/**
+ * Describes the message jammi.v1.training.FineTuneSpec.
+ * Use `create(FineTuneSpecSchema)` to create a new message.
+ */
+export declare const FineTuneSpecSchema: GenMessage<FineTuneSpec>;
+/**
+ * A graph-supervised fine-tune. Mirrors the engine's `TrainingSpec::GraphFineTune`.
+ *
+ * @generated from message jammi.v1.training.GraphFineTuneSpec
+ */
+export type GraphFineTuneSpec = Message<"jammi.v1.training.GraphFineTuneSpec"> & {
+    /**
+     * Node-text + edge sources and their column bindings.
+     *
+     * @generated from field: jammi.v1.training.GraphFineTuneSources sources = 1;
+     */
+    sources?: GraphFineTuneSources | undefined;
+    /**
+     * Walk / negative-sampling knobs, including the deterministic seed.
+     *
+     * @generated from field: jammi.v1.training.GraphSampleConfig sample_config = 2;
+     */
+    sampleConfig?: GraphSampleConfig | undefined;
+};
+/**
+ * Describes the message jammi.v1.training.GraphFineTuneSpec.
+ * Use `create(GraphFineTuneSpecSchema)` to create a new message.
+ */
+export declare const GraphFineTuneSpecSchema: GenMessage<GraphFineTuneSpec>;
+/**
+ * An episodic in-context-predictor meta-training. Mirrors the engine's
+ * `TrainingSpec::ContextPredictor`. The predictor carries its full budget inside
+ * `predictor_spec`, so this variant needs no common base-model / config block.
+ *
+ * @generated from message jammi.v1.training.ContextPredictorSpec
+ */
+export type ContextPredictorSpec = Message<"jammi.v1.training.ContextPredictorSpec"> & {
+    /**
+     * Source whose embedding table the episodic context is sampled from.
+     *
+     * @generated from field: string source = 1;
+     */
+    source: string;
+    /**
+     * The architecture / objective / episodic / optimisation specification.
+     *
+     * @generated from field: jammi.v1.training.ContextPredictorTrainConfig predictor_spec = 2;
+     */
+    predictorSpec?: ContextPredictorTrainConfig | undefined;
+};
+/**
+ * Describes the message jammi.v1.training.ContextPredictorSpec.
+ * Use `create(ContextPredictorSpecSchema)` to create a new message.
+ */
+export declare const ContextPredictorSpecSchema: GenMessage<ContextPredictorSpec>;
+/**
+ * @generated from message jammi.v1.training.StartTrainingRequest
+ */
+export type StartTrainingRequest = Message<"jammi.v1.training.StartTrainingRequest"> & {
+    /**
+     * The self-describing training spec — the verb that produces the job. Mirrors
+     * the engine's `TrainingSpec` enum: the variant selects the submit path, and
+     * its fields carry everything a worker reconstructs the run from.
+     *
+     * @generated from oneof jammi.v1.training.StartTrainingRequest.spec
+     */
+    spec: {
+        /**
+         * @generated from field: jammi.v1.training.FineTuneSpec fine_tune = 1;
+         */
+        value: FineTuneSpec;
+        case: "fineTune";
+    } | {
+        /**
+         * @generated from field: jammi.v1.training.GraphFineTuneSpec graph_fine_tune = 2;
+         */
+        value: GraphFineTuneSpec;
+        case: "graphFineTune";
+    } | {
+        /**
+         * @generated from field: jammi.v1.training.ContextPredictorSpec context_predictor = 3;
+         */
+        value: ContextPredictorSpec;
+        case: "contextPredictor";
+    } | {
+        case: undefined;
+        value?: undefined;
+    };
+    /**
+     * Base encoder to adapt for the two LoRA fine-tune kinds (`local:<path>`, an
+     * HF repo id, or a fine-tuned id). Unused by the context-predictor kind (it
+     * names its source's embedding model implicitly), so it is empty there.
+     *
+     * @generated from field: string base_model = 4;
+     */
+    baseModel: string;
+    /**
+     * Optional fine-tune hyperparameters for the two LoRA fine-tune kinds; unset
+     * applies the engine's defaults. Unused by the context-predictor kind (its
+     * budget rides inside `predictor_spec`).
+     *
+     * @generated from field: optional jammi.v1.training.FineTuneConfig config = 5;
+     */
+    config?: FineTuneConfig | undefined;
+};
+/**
+ * Describes the message jammi.v1.training.StartTrainingRequest.
+ * Use `create(StartTrainingRequestSchema)` to create a new message.
+ */
+export declare const StartTrainingRequestSchema: GenMessage<StartTrainingRequest>;
+/**
+ * @generated from message jammi.v1.training.StartTrainingResponse
+ */
+export type StartTrainingResponse = Message<"jammi.v1.training.StartTrainingResponse"> & {
+    /**
+     * Server-assigned job id. Poll `TrainingStatus` with it.
+     *
+     * @generated from field: string job_id = 1;
+     */
+    jobId: string;
+    /**
+     * The deterministic output model id the trained artifact registers under —
+     * the same id the engine's `TrainingJob::model_id` carries at submit time.
+     *
+     * @generated from field: string model_id = 2;
+     */
+    modelId: string;
+};
+/**
+ * Describes the message jammi.v1.training.StartTrainingResponse.
+ * Use `create(StartTrainingResponseSchema)` to create a new message.
+ */
+export declare const StartTrainingResponseSchema: GenMessage<StartTrainingResponse>;
+/**
+ * @generated from message jammi.v1.training.TrainingStatusRequest
+ */
+export type TrainingStatusRequest = Message<"jammi.v1.training.TrainingStatusRequest"> & {
+    /**
+     * Job id returned by a prior StartTraining call.
+     *
+     * @generated from field: string job_id = 1;
+     */
+    jobId: string;
+};
+/**
+ * Describes the message jammi.v1.training.TrainingStatusRequest.
+ * Use `create(TrainingStatusRequestSchema)` to create a new message.
+ */
+export declare const TrainingStatusRequestSchema: GenMessage<TrainingStatusRequest>;
+/**
+ * @generated from message jammi.v1.training.TrainingStatusResponse
+ */
+export type TrainingStatusResponse = Message<"jammi.v1.training.TrainingStatusResponse"> & {
+    /**
+     * Current lifecycle status: "queued", "running", "completed", or "failed".
+     *
+     * @generated from field: string status = 1;
+     */
+    status: string;
+    /**
+     * The output model id the trained artifact registers under.
+     *
+     * @generated from field: string model_id = 2;
+     */
+    modelId: string;
+    /**
+     * The failure message when `status` is "failed"; empty otherwise.
+     *
+     * @generated from field: string error = 3;
+     */
+    error: string;
+};
+/**
+ * Describes the message jammi.v1.training.TrainingStatusResponse.
+ * Use `create(TrainingStatusResponseSchema)` to create a new message.
+ */
+export declare const TrainingStatusResponseSchema: GenMessage<TrainingStatusResponse>;
+/**
+ * Fine-tuning method. Mirrors the engine's `FineTuneMethod`; LoRA is the only
+ * method, trained as either an encoder-internal adapter or a projection head
+ * over a frozen base model depending on `FineTuneConfig.target_modules`.
+ *
+ * @generated from enum jammi.v1.training.FineTuneMethod
+ */
+export declare enum FineTuneMethod {
+    /**
+     * @generated from enum value: FINE_TUNE_METHOD_UNSPECIFIED = 0;
+     */
+    FINE_TUNE_METHOD_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: LORA = 1;
+     */
+    LORA = 1
+}
+/**
+ * Describes the enum jammi.v1.training.FineTuneMethod.
+ */
+export declare const FineTuneMethodSchema: GenEnum<FineTuneMethod>;
+/**
+ * Classification-loss function. Mirrors the engine's `ClassificationLoss`.
+ *
+ * @generated from enum jammi.v1.training.ClassificationLoss
+ */
+export declare enum ClassificationLoss {
+    /**
+     * @generated from enum value: CLASSIFICATION_LOSS_UNSPECIFIED = 0;
+     */
+    CLASSIFICATION_LOSS_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: CROSS_ENTROPY = 1;
+     */
+    CROSS_ENTROPY = 1
+}
+/**
+ * Describes the enum jammi.v1.training.ClassificationLoss.
+ */
+export declare const ClassificationLossSchema: GenEnum<ClassificationLoss>;
+/**
+ * Which loss signal drives early stopping. Mirrors `EarlyStoppingMetric`.
+ *
+ * @generated from enum jammi.v1.training.EarlyStoppingMetric
+ */
+export declare enum EarlyStoppingMetric {
+    /**
+     * @generated from enum value: EARLY_STOPPING_METRIC_UNSPECIFIED = 0;
+     */
+    EARLY_STOPPING_METRIC_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: VAL_LOSS = 1;
+     */
+    VAL_LOSS = 1,
+    /**
+     * @generated from enum value: TRAIN_LOSS = 2;
+     */
+    TRAIN_LOSS = 2
+}
+/**
+ * Describes the enum jammi.v1.training.EarlyStoppingMetric.
+ */
+export declare const EarlyStoppingMetricSchema: GenEnum<EarlyStoppingMetric>;
+/**
+ * Learning-rate schedule after warmup. Mirrors the engine's `LrSchedule`.
+ *
+ * @generated from enum jammi.v1.training.LrSchedule
+ */
+export declare enum LrSchedule {
+    /**
+     * @generated from enum value: LR_SCHEDULE_UNSPECIFIED = 0;
+     */
+    LR_SCHEDULE_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: CONSTANT = 1;
+     */
+    CONSTANT = 1,
+    /**
+     * @generated from enum value: COSINE_DECAY = 2;
+     */
+    COSINE_DECAY = 2,
+    /**
+     * @generated from enum value: LINEAR_DECAY = 3;
+     */
+    LINEAR_DECAY = 3
+}
+/**
+ * Describes the enum jammi.v1.training.LrSchedule.
+ */
+export declare const LrScheduleSchema: GenEnum<LrSchedule>;
+/**
+ * LoRA A/B matrix initialisation. Mirrors `jammi_lora::LoraInitMode`.
+ *
+ * @generated from enum jammi.v1.training.LoraInitMode
+ */
+export declare enum LoraInitMode {
+    /**
+     * @generated from enum value: LORA_INIT_MODE_UNSPECIFIED = 0;
+     */
+    LORA_INIT_MODE_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: ZEROS_B = 1;
+     */
+    ZEROS_B = 1,
+    /**
+     * @generated from enum value: GAUSSIAN = 2;
+     */
+    GAUSSIAN = 2
+}
+/**
+ * Describes the enum jammi.v1.training.LoraInitMode.
+ */
+export declare const LoraInitModeSchema: GenEnum<LoraInitMode>;
+/**
+ * Dtype for the frozen backbone weights. Mirrors `jammi_lora::BackboneDtype`.
+ *
+ * @generated from enum jammi.v1.training.BackboneDtype
+ */
+export declare enum BackboneDtype {
+    /**
+     * @generated from enum value: BACKBONE_DTYPE_UNSPECIFIED = 0;
+     */
+    BACKBONE_DTYPE_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: F32 = 1;
+     */
+    F32 = 1,
+    /**
+     * @generated from enum value: BF16 = 2;
+     */
+    BF16 = 2,
+    /**
+     * @generated from enum value: F16 = 3;
+     */
+    F16 = 3
+}
+/**
+ * Describes the enum jammi.v1.training.BackboneDtype.
+ */
+export declare const BackboneDtypeSchema: GenEnum<BackboneDtype>;
+/**
+ * Edge provenance for a graph fine-tune. Mirrors the engine's `EdgeProvenance`
+ * — the load-bearing circularity distinction (declared edges teach the metric
+ * something new; similarity edges largely re-learn it).
+ *
+ * @generated from enum jammi.v1.training.EdgeProvenance
+ */
+export declare enum EdgeProvenance {
+    /**
+     * @generated from enum value: EDGE_PROVENANCE_UNSPECIFIED = 0;
+     */
+    EDGE_PROVENANCE_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: DECLARED = 1;
+     */
+    DECLARED = 1,
+    /**
+     * @generated from enum value: SIMILARITY = 2;
+     */
+    SIMILARITY = 2
+}
+/**
+ * Describes the enum jammi.v1.training.EdgeProvenance.
+ */
+export declare const EdgeProvenanceSchema: GenEnum<EdgeProvenance>;
+/**
+ * Which curated in-context-predictor architecture to train. Mirrors the
+ * engine's `ContextArchitecture`.
+ *
+ * @generated from enum jammi.v1.training.ContextArchitecture
+ */
+export declare enum ContextArchitecture {
+    /**
+     * @generated from enum value: CONTEXT_ARCHITECTURE_UNSPECIFIED = 0;
+     */
+    CONTEXT_ARCHITECTURE_UNSPECIFIED = 0,
+    /**
+     * @generated from enum value: CNP = 1;
+     */
+    CNP = 1,
+    /**
+     * @generated from enum value: ATTN_CNP = 2;
+     */
+    ATTN_CNP = 2,
+    /**
+     * @generated from enum value: TNP = 3;
+     */
+    TNP = 3
+}
+/**
+ * Describes the enum jammi.v1.training.ContextArchitecture.
+ */
+export declare const ContextArchitectureSchema: GenEnum<ContextArchitecture>;
+/**
+ * TrainingService — the wire surface for the engine's training verbs.
+ *
+ * `StartTraining` and `TrainingStatus` are imperative session methods, not SQL
+ * queries, so they live as typed gRPC RPCs alongside SessionService /
+ * TriggerService / EmbeddingService / InferenceService / EvalService rather
+ * than on the Flight SQL surface (which carries query/result only, per ADR-01
+ * §3.2). Training is a substrate primitive — meta-learning an adapter or a
+ * predictor over a registered source — and names no tenant-domain concept.
+ *
+ * The service serves all three engine training kinds (`fine_tune`,
+ * `graph_fine_tune`, `context_predictor`) behind one verb: `StartTraining`
+ * carries a full `TrainingSpec` oneof and dispatches to the matching engine
+ * submit verb, returning the durable job id and the (deterministic) output
+ * model id. The job runs asynchronously inside the engine; the client polls
+ * `TrainingStatus` by id until a terminal state. There is no progress stream —
+ * the abstraction exposes none, so this surface invents none. The service
+ * reimplements no training, LoRA, sampling, or scan logic; it is purely a wire
+ * adapter.
+ *
+ * Tenant scope is read from the request's `SessionTenant` extension, set by the
+ * shared TenantInterceptor from the `jammi-session-id` header — the same
+ * mechanism every other gRPC service and Flight SQL already use.
+ *
+ * @generated from service jammi.v1.training.TrainingService
+ */
+export declare const TrainingService: GenService<{
+    /**
+     * Submit a training job over a registered source and return its id + output
+     * model id. The job trains asynchronously; poll `TrainingStatus` with the
+     * returned job id. Peer of the engine's `fine_tune` / `fine_tune_graph` /
+     * `train_context_predictor` submit verbs (the spec oneof selects which).
+     *
+     * @generated from rpc jammi.v1.training.TrainingService.StartTraining
+     */
+    startTraining: {
+        methodKind: "unary";
+        input: typeof StartTrainingRequestSchema;
+        output: typeof StartTrainingResponseSchema;
+    };
+    /**
+     * Read a training job's current status by id, plus its output model id and
+     * the failure message when it failed. Peer of reading back the engine's
+     * `TrainingJob` handle — the output model id and error are what let a remote
+     * `wait()` retrieve the result and a failure reason.
+     *
+     * @generated from rpc jammi.v1.training.TrainingService.TrainingStatus
+     */
+    trainingStatus: {
+        methodKind: "unary";
+        input: typeof TrainingStatusRequestSchema;
+        output: typeof TrainingStatusResponseSchema;
+    };
+}>;
+//# sourceMappingURL=training_pb.d.ts.map