@zhuoyuezs/ml-platform 0.2.2-alpha.2 → 0.2.2-alpha.3

package/checksums.json

@@ -2,8 +2,8 @@
   "files": [
     {
       "path": "runtime/business-client/README.md",
-      "sha256": "sha256:7e84fad64f8ee6c7aae467b7947123f3a490b799ec5ffa24fd008e0fa37b133e",
-      "size_bytes": 7898
+      "sha256": "sha256:30d921fabf5bc0339eda60169a1865f6007807b1d9a8d3e646b42f68cd563240",
+      "size_bytes": 8125
     },
     {
       "path": "runtime/business-client/package-lock.json",
@@ -142,8 +142,8 @@
     },
     {
       "path": "skills/model-lifecycle-management/SKILL.md",
-      "sha256": "sha256:06307c24ee24ef28978183c9c83a5542fa5b21ec9912e717e84bcede80434cd8",
-      "size_bytes": 4049
+      "sha256": "sha256:4ea40815d73b3a470b2572874dcad63640cbdd53d3c10293d7a5ce8621a13b96",
+      "size_bytes": 4272
     },
     {
       "path": "skills/model-lifecycle-management/agents/openai.yaml",
@@ -152,8 +152,8 @@
     },
     {
       "path": "skills/model-lifecycle-management/references/discovery.md",
-      "sha256": "sha256:b2560e3c79fd309ce15a45d6bc3c30f2206d71f88197403cfa432580f9e4c50e",
-      "size_bytes": 5075
+      "sha256": "sha256:184472a79fb217dd8c74dffab20e75285b0c8f69207dc593ae3099b8a2b6d6d5",
+      "size_bytes": 5381
     },
     {
       "path": "skills/model-lifecycle-management/references/evaluation.md",
@@ -167,13 +167,13 @@
     },
     {
       "path": "skills/model-lifecycle-management/references/training-contracts.md",
-      "sha256": "sha256:06af04539ab8d69467680e4a662595719a40f82e23fac19b9d4460be189201c5",
-      "size_bytes": 12887
+      "sha256": "sha256:4cecf5f90464640d063ca721fded1c27b46bb79daa0f67c78a8fb6b0e3e60b3b",
+      "size_bytes": 14747
     },
     {
       "path": "skills/model-lifecycle-management/references/training.md",
-      "sha256": "sha256:f754b0a7717914b2ac028feac5b1b9ef59ff673e7ffe722f37f5ae3d5e6e0bfd",
-      "size_bytes": 5686
+      "sha256": "sha256:97de7894da24a07b4146d6de5e00a0cf17166654fb219a145f35adff5e8a7f99",
+      "size_bytes": 7980
     }
   ],
   "schema_version": "data_platform.ml_platform_checksums/v1"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhuoyuezs/ml-platform",
-  "version": "0.2.2-alpha.2",
+  "version": "0.2.2-alpha.3",
   "description": "ML 数据平台 business CLI 与 Agent Skills 管理工具",
   "bin": {
     "ml-platform": "scripts/main.js"

package/release.json

@@ -13,7 +13,7 @@
     "entrypoint": "src/cli.js",
     "name": "ml-platform",
     "path": "runtime/business-client",
-    "sha256": "sha256:b439dd6ae614439aca3e812f6e2127f743a7efaf4ffa090886fdebed55c9824d",
+    "sha256": "sha256:ceb3fdabf57044a5bf8d54baf04ab10327cb896e4eaff3897c18e08e96e5336a",
     "version": "0.8.2"
   },
   "policy_sha256": "sha256:02fbde0134696b78c0c365d6e0b76c9814227208da5b2b3d4e5dbd82dc4a3606",
@@ -25,7 +25,7 @@
     "ml_data_platform.feature_set/v1",
     "ml_data_platform.dataset_manifest/v1"
   ],
-  "release_version": "0.2.2-alpha.2",
+  "release_version": "0.2.2-alpha.3",
   "runtime_requirements": {
     "node": ">=18",
     "os": [
@@ -38,21 +38,21 @@
     "feature-management": {
       "path": "skills/feature-management",
       "requires_cli": ">=0.8.2 <0.9.0",
-      "revision": "0.2.2-alpha.2",
+      "revision": "0.2.2-alpha.3",
       "sha256": "sha256:d32f0306f129d7bf76eb037fe239c84dd3c29c64b61600799d6e40320784bb10"
     },
     "model-deployment-management": {
       "path": "skills/model-deployment-management",
       "requires_cli": ">=0.8.2 <0.9.0",
-      "revision": "0.2.2-alpha.2",
+      "revision": "0.2.2-alpha.3",
       "sha256": "sha256:e7d87c272a0dcc028ea628d5e36f5b7e52ed27cb9c1bc17ea76527478e28b46d"
     },
     "model-lifecycle-management": {
       "path": "skills/model-lifecycle-management",
       "requires_cli": ">=0.8.2 <0.9.0",
-      "revision": "0.2.2-alpha.2",
-      "sha256": "sha256:f841821d2ca90198ad07236953a864393cb1e8db51cb349f0ded85190ebc1c10"
+      "revision": "0.2.2-alpha.3",
+      "sha256": "sha256:ec6ebb2587fd7acd024def75a8c0e9cd94b6cdd8ee6df439609291b496751b91"
     }
   },
-  "source_commit": "378de56c68d4ea525083170cd500eeef01e88bd1"
+  "source_commit": "e713953153439c28680b60b56f57d3b9679ff49a"
 }

package/runtime/business-client/README.md

@@ -147,6 +147,9 @@ resolve the same contracts as submission without registering or scheduling
 workloads. Read `submitted=false` and `unchecked`: successful preflight does not
 verify image pulls, cluster resources or runtime execution. Complete request
 examples and identity propagation are in the bundled model lifecycle Skill.
+Training Run and Job validation also return `effective_trainer_parameters`;
+for Job preflight it must match `job_spec.trainer_parameters` and represents
+the complete mapping delivered to the Worker after defaults are resolved.
 
 Maintainers can export the current API, including uncommitted schema changes,
 with `uv run python scripts/generate_openapi_schema.py --working-tree` from the

package/skills/model-lifecycle-management/SKILL.md

@@ -24,10 +24,12 @@ Before training a model that will be evaluated, also read
 [references/evaluation.md](references/evaluation.md). Check the final dataset
 schema, training split binding and evaluation runtime before submission.
 OpenAPI free-form dictionaries are not complete nested contracts: use the
-examples in these references and same-release validation, not successive 422s
-to discover fields. Preserve release/error evidence when deployment differs.
+SystemModel `parameter_schema`, examples in these references and same-release
+validation, not successive 422s to discover fields. Preserve release/error
+evidence when deployment differs.
 
 - Establish exact immutable inputs, canonical SystemModel `v1`, runtime-release image digest, model inputs/targets, resources, evaluation policy, and package destination from authoritative contracts. Do not infer them from model names.
+- For Chronos-2, inspect the typed `lora_config` schema before authoring a Run and preserve `effective_trainer_parameters` from validation. Do not infer PEFT keys or rely on Worker defaults.
 - Validate, register, submit, retry, cancel, and package creation are separate actions. Read-only discovery and validation do not authorize mutation or workload submission.
 - Before submission, show the exact JSON path, immutable identities, expected workload, and target API. Obtain explicit authorization. Retry and cancel require separate authorization for the exact Job.
 - Never use `latest` image tags. During development the canonical SystemModel stays at `v1`; image changes are reviewed Runtime Release updates and require confirming no active Job depends on the previous binding.

package/skills/model-lifecycle-management/references/discovery.md

@@ -68,6 +68,10 @@ artifact eligibility, selected runtime, input modes and immutable-version confli
 checks used by creation. Both return `status=valid`, the proposed `job_spec` or
 `package`, `submitted=false` and `unchecked` execution checks. They do not register
 Jobs/packages or invoke schedulers. Artifact reads can populate local caches.
+For training, confirm that top-level `effective_trainer_parameters` exactly
+matches `job_spec.trainer_parameters`; this is the complete parameter mapping
+the Worker receives after platform defaults and reproducibility controls are
+resolved.
 
 A valid preview does not prove image availability, cluster capacity, credentials,
 framework imports, training success or packaging test success. Read `unchecked`
@@ -81,9 +85,11 @@ Training Run `task`, `data`, `data.artifact`, `features.feature_set`, `splits`,
 `temporal` and `trainer` are explicit objects in OpenAPI. Evaluation artifact,
 config references, members and execution, plus ModelPackage requests, are also
 structured. Unknown keys return 422 with `detail[].loc` identifying the field.
-Keep algorithm-specific `trainer.parameters` and documented plugin configuration
-maps extensible; validate them against the SystemModel parameter schema. Do not treat
-all remaining `additionalProperties` as a platform schema defect.
+Keep algorithm-specific `trainer.parameters` governed by the SystemModel
+parameter schema. A typed nested contract such as Chronos-2 `lora_config`
+rejects unknown fields; a SystemModel that intentionally advertises
+`additionalProperties` remains extensible. Do not infer one policy from the
+other.
 
 These HTTP checks preserve accepted values and do not rewrite historical stored
 contracts or hashes. Legacy records remain readable. A previously ignored typo

package/skills/model-lifecycle-management/references/training-contracts.md

@@ -68,6 +68,59 @@ Non-temporal regression can omit it. Do not change task kind just to evade check
 `series_keys` must identify entity-key columns when multiple series coexist;
 an empty list treats the entire artifact as one series.
 
+## Chronos-2 LoRA parameters
+
+Chronos-2 fine-tuning settings belong under `trainer.parameters`; they are
+frozen with the immutable TrainingRun. This fragment overrides rank and keeps
+the other released LoRA defaults explicit:
+
+```json
+{
+  "trainer": {
+    "project": "default",
+    "name": "chronos2_forecaster",
+    "version": "v1",
+    "parameters": {
+      "finetune_mode": "lora",
+      "num_steps": 1000,
+      "batch_size": 16,
+      "learning_rate": 0.00001,
+      "gradient_accumulation_steps": 2,
+      "logging_steps": 10,
+      "lora_config": {
+        "r": 16,
+        "lora_alpha": 16,
+        "lora_dropout": 0.0,
+        "target_modules": [
+          "self_attention.q",
+          "self_attention.v",
+          "self_attention.k",
+          "self_attention.o",
+          "output_patch_embedding.output_layer"
+        ]
+      }
+    }
+  },
+  "reproducibility": {"seed": 42, "deterministic": true}
+}
+```
+
+Always obtain the accepted fields, target-module enum and defaults from
+`get-system-model chronos2_forecaster v1`; unknown nested LoRA keys are
+rejected before scheduling. `r >= 1`, `lora_alpha >= 1`, and
+`0 <= lora_dropout < 1`. `finetune_mode=full` rejects `lora_config`.
+
+The platform materializes omitted values before Job creation. Run validation
+returns `effective_trainer_parameters`; Job validation returns the same object
+both at the top level and in `job_spec.trainer_parameters`. Review that object
+instead of assuming library defaults. Changing any Trainer parameter requires
+a new TrainingRun identity; Job retry cannot alter it.
+
+Do not put `prediction_length` in this mapping. It is derived from the temporal
+horizon and frequency. Trainer `context_length` defaults to the sequence input
+adapter's context length, and `min_past` defaults to prediction length. GPU,
+CPU, memory and deadline remain TrainingJob execution fields.
+
 ## Direct tabular forecasting with known-future features
 
 Tree and other matrix-based models can forecast a variable target grid without

package/skills/model-lifecycle-management/references/training.md

@@ -85,6 +85,70 @@ ml-platform --profile server cancel-training-job <job-id>
 
 Retry and cancellation require exact-Job authorization. Never change a run contract while retrying; create a new immutable run when semantic inputs change.
 
+## Chronos-2 fine-tuning parameters
+
+Configure Chronos-2 algorithm parameters only under
+`trainer.parameters` in the TrainingRun. Configure execution device and GPU
+resources later under TrainingJob `execution`; do not mix the two contracts.
+The Run is immutable after registration, so changing rank, learning rate,
+batch size or fine-tuning mode requires a new `run_id`. Retrying an existing
+Job must retain the original parameters.
+
+Inspect the same-release contract first:
+
+```bash
+ml-platform --profile server get-system-model \
+  chronos2_forecaster v1 --project default
+```
+
+The returned `parameter_schema` is authoritative. A typical explicit LoRA
+fragment is:
+
+```json
+{
+  "trainer": {
+    "project": "default",
+    "name": "chronos2_forecaster",
+    "version": "v1",
+    "parameters": {
+      "finetune_mode": "lora",
+      "num_steps": 1000,
+      "batch_size": 16,
+      "learning_rate": 0.00001,
+      "gradient_accumulation_steps": 2,
+      "logging_steps": 10,
+      "lora_config": {
+        "r": 16,
+        "lora_alpha": 32,
+        "lora_dropout": 0.05
+      }
+    }
+  },
+  "reproducibility": {"seed": 42, "deterministic": true}
+}
+```
+
+`r` and `lora_alpha` must be positive integers; `lora_dropout` is in `[0, 1)`.
+Only target modules enumerated by the SystemModel schema are accepted. LoRA is
+the default mode; full fine-tuning rejects `lora_config`. Omitted LoRA settings
+resolve to `r=8`, `lora_alpha=16`, `lora_dropout=0` and the released Chronos-2
+target-module list. Omitted learning rate resolves to `1e-5` for LoRA and
+`1e-6` for full fine-tuning. `min_past` defaults to the derived prediction
+length, while trainer `context_length` defaults to
+`input_adapter.context_length`.
+
+Set random seed and determinism only through top-level `reproducibility`.
+Compatibility fields with the same names may appear in the parameter schema,
+but the top-level values are authoritative. Do not configure
+`prediction_length`; the platform derives it as `temporal.horizon /
+temporal.frequency`.
+
+After `validate-training-run`, inspect and preserve
+`effective_trainer_parameters`. After Run registration, call
+`validate-training-job` and confirm that its top-level
+`effective_trainer_parameters` exactly equals
+`job_spec.trainer_parameters` before requesting submission authorization.
+
 ## Chronos-2 future-known inputs
 
 For storage forecasting, tomorrow's weather forecast and peak/valley schedule