fix(dataset): align validation rules with platform data format spec

- Support content array format [{text/image/video}] alongside legacy string
- Add tool role support with tool_calls structure and tool_call_id validation
- Add thinking tag placement check (only in last assistant message)
- Add OpenAI migration guards: warn on unsupported name/weight fields
- Add loss_weight range validation (0.0–1.0)
- Fix size limits: SFT/DPO 200MB, CPT 300MB, media ZIP 2GB
- Enforce data.jsonl at ZIP root (reject nested wrapping folders)
- Add ZIP filename constraints: charset [a-zA-Z0-9_-], length ≤120, uniqueness
- Add .tif to accepted image extensions
- Add DPO_LAST_MSG_NOT_USER warning when messages don't end with user role
- CPT profile now uses dedicated 300MB cap instead of shared default
- Expand unit tests from 19 to 45 covering all new validation paths

Author: gujieye

packages/commands/src/commands/dataset/upload.ts

@@ -5,6 +5,7 @@ import {
   parseDatasetSchemaFlag,
   formatIssue,
   MAX_DATASET_BYTES,
+  MAX_CPT_BYTES,
   MAX_MEDIA_ZIP_BYTES,
   BailianError,
   ExitCode,
@@ -16,7 +17,7 @@ const UPLOAD_FLAGS = {
   file: {
     type: "string",
     valueHint: "<path>",
-    description: "Local dataset file (.jsonl or .zip; ≤300MB text, ≤1GB image)",
+    description: "Local dataset file (.jsonl or .zip; ≤200MB SFT/DPO, ≤300MB CPT, ≤2GB media zip)",
     required: true,
   },
   purpose: {
@@ -57,13 +58,14 @@ export default defineCommand({
   ],
   notes: [
     "Supports .jsonl (text) and .zip (audio/image archives with a data.jsonl",
-    "manifest). Five record schemas are recognized: chatml = {messages:[...]}",
+    "manifest). Six record schemas are recognized: chatml = {messages:[...]}",
     '(SFT); dpo = {messages:[...], chosen, rejected}; cpt = {text:"..."}',
     '(continual pre-training, raw text); tts = {wav_fn:"train/xxx.wav",',
     'text:"..."} (audio fine-tuning); image = {img_path:"..."} (image',
-    "generation). With no --schema, a record carrying wav_fn is validated as",
-    "TTS, img_path as image, chosen/rejected as DPO, text (no messages) as CPT,",
-    "otherwise ChatML. Upload cap: 300MB text, 1GB image. Upload uses the",
+    "generation); video = {first_frame_path:...} (video generation). With no",
+    "--schema, a record carrying wav_fn is validated as TTS, img_path as image,",
+    "chosen/rejected as DPO, text (no messages) as CPT, otherwise ChatML.",
+    "Upload cap: 200MB SFT/DPO text, 300MB CPT, 2GB media zip. Upload uses the",
     "OpenAI-compatible /compatible-mode/v1/files endpoint so the purpose tag is",
     "persisted (the DashScope-native /api/v1/files drops it).",
   ],
@@ -72,11 +74,15 @@ export default defineCommand({
     const filePath = flags.file;
     const purpose = flags.purpose || "fine-tune";
     const schema = parseDatasetSchemaFlag(flags.schema);
-    // Image and video schemas allow larger ZIPs (1 GB vs 300 MB for text).
+    // Size caps differ per training type: SFT/DPO 200MB, CPT 300MB, media ZIP 2GB.
     const isMediaSchema = schema === "image" || schema === "video";
+    const maxBytes = isMediaSchema
+      ? MAX_MEDIA_ZIP_BYTES
+      : schema === "cpt"
+        ? MAX_CPT_BYTES
+        : MAX_DATASET_BYTES;
 
     if (!flags.noValidate) {
-      const maxBytes = isMediaSchema ? MAX_MEDIA_ZIP_BYTES : MAX_DATASET_BYTES;
       const result = await validateDataset(filePath, {
         fullValidate: flags.fullValidate,
         schema,
@@ -116,7 +122,7 @@ export default defineCommand({
           action: "dataset.upload",
           file: filePath,
           purpose,
-          max_bytes: isMediaSchema ? MAX_MEDIA_ZIP_BYTES : MAX_DATASET_BYTES,
+          max_bytes: maxBytes,
           validate: !flags.noValidate,
           schema: schema ?? "auto",
         },

packages/core/src/dataset/index.ts

@@ -7,6 +7,7 @@ export {
   registerValidator,
   listSupportedFormats,
   MAX_DATASET_BYTES,
+  MAX_CPT_BYTES,
   MAX_MEDIA_ZIP_BYTES,
   parseDatasetSchemaFlag,
   formatIssue,

packages/core/src/dataset/validate/common.ts

@@ -12,18 +12,24 @@ import { ExitCode } from "../../errors/codes.ts";
 import type { DatasetSchema, ValidationIssue, ValidationStats } from "./types.ts";
 
 /**
- * The platform caps dataset uploads at 300MB per file. `bl dataset upload`
- * enforces this client-side so users learn early. Update if the platform
- * raises the cap or differentiates per-purpose limits.
+ * The platform caps SFT/DPO text dataset uploads at 200MB per file.
+ * `bl dataset upload` enforces this client-side so users learn early.
+ * CPT uses 300MB (see MAX_CPT_BYTES); API general upload is also 300MB.
  */
-export const MAX_DATASET_BYTES = 300 * 1024 * 1024;
+export const MAX_DATASET_BYTES = 200 * 1024 * 1024;
 
 /**
- * Image / video ZIP size cap — 1 GB per the platform docs (vs 300 MB for
- * text / audio). Used by `bl dataset upload` for media schemas and by the
- * `sft-lora` training profile for image / video validation.
+ * CPT text dataset size cap — 300 MB per the platform docs.
+ * CPT requires at least 50M tokens; larger files are expected.
  */
-export const MAX_MEDIA_ZIP_BYTES = 1024 * 1024 * 1024;
+export const MAX_CPT_BYTES = 300 * 1024 * 1024;
+
+/**
+ * Image / video ZIP size cap — 2 GB per the platform docs. Used by
+ * `bl dataset upload` for media schemas and by the `sft-lora` training
+ * profile for image / video validation.
+ */
+export const MAX_MEDIA_ZIP_BYTES = 2 * 1024 * 1024 * 1024;
 
 export interface PreflightResult {
   bytes: number;

packages/core/src/dataset/validate/index.ts

@@ -4,7 +4,12 @@ export {
   registerValidator,
   listSupportedFormats,
 } from "./registry.ts";
-export { MAX_DATASET_BYTES, MAX_MEDIA_ZIP_BYTES, parseDatasetSchemaFlag } from "./common.ts";
+export {
+  MAX_DATASET_BYTES,
+  MAX_CPT_BYTES,
+  MAX_MEDIA_ZIP_BYTES,
+  parseDatasetSchemaFlag,
+} from "./common.ts";
 export { formatIssue } from "./format.ts";
 export type {
   ValidatorSpec,