diff --git a/openspec/changes/ts-as-modality/tasks.md b/openspec/changes/ts-as-modality/tasks.md
index 6788b6a..74be003 100644
--- a/openspec/changes/ts-as-modality/tasks.md
+++ b/openspec/changes/ts-as-modality/tasks.md
@@ -19,7 +19,7 @@
 - [x] 2.2 Projector `model/projector.py`：`Linear(256→896)+LayerNorm`（验证：输出 `[B,n_patches,896]` 对齐 Qwen2.5-0.5B hidden）
 - [x] 2.3 多模态拼接 `model/multimodal.py`：Qwen tokenizer tokenize 文本部分，TS token 作 inputs_embeds 插入 `[属性][时间戳][TS tok][事件][问题][回答][EOS]`，统一构造 inputs_embeds+attention_mask+labels（仅回答段非 -100）（验证：8 项单测过——seq_len ≤1024、mask 全 1、回答段 label 非 -100、TS token 计入序列、无 NaN、超长左截断）。spec 澄清：训练拼回答+EOS，用 `len(tokenizer)` 构建嵌入表以容纳 special tokens。
 - [x] 2.4 训练/推理封装 `model/wrapper.py`：`MultimodalTSModel` 组合 Encoder+Projector+LLM+LoRA 挂载开关，`forward` 返 loss、`generate` 返文本，支持 `freeze_llm`(阶段①)/`enable_lora(r=16)`(阶段②)（验证：端到端 forward+backward 跑通、grad 流入 Encoder/Projector、generate 出文本、阶段①峰值 1.87GB ≪5GB 设计预算）。spec 澄清：wrapper 既接收原始 batch（series+文本列表）走端到端，也兼容预拼接 inputs_embeds；encoder/projector fp32，输出投影到 LLM dtype(bf16)。
-- [ ] 2.5 Collator `data/collator.py`：JSONL→batch 张量，处理变长 C（padding+mask）与变长文本（padding+attention_mask）（验证：batch=4 张量形状一致无 NaN）
+- [x] 2.5 Collator `data/collator.py`：JSONL→wrapper 原始 batch（series 张量 + attributes/timestamps/events/question/answer 文本列表），处理变长 T（padding/trunc 到 max_T）与变长 C（padding 到 max_C）、NaN→fill；变长文本 padding/attention_mask 由 wrapper 逐样本 splice+max_seq 填充处理（验证：11 项单测过；端到端 Collator→Wrapper fwd+bwd+generate 跑通）
 - [ ] 2.6 M2 出口验证：单 batch 前向+反向在 3060 跑通，loss 有限且下降趋势，阶段①模式峰值显存 ~5GB
 
 ## 3. M3 · 阶段①对齐训练
diff --git a/src/tsmm/data/collator.py b/src/tsmm/data/collator.py
new file mode 100644
index 0000000..0dff93f
--- /dev/null
+++ b/src/tsmm/data/collator.py
@@ -0,0 +1,103 @@
+"""Collator (T2.5): JSONL sample dicts → wrapper raw batch.
+
+Input  : a list of sample dicts (the gen_synthetic / real_bench JSONL schema):
+    {
+      "series":     [[T, C] floats],          # may contain NaN (missing)
+      "attributes": [{"name","unit","freq"}],
+      "timestamps": [T ints],
+      "events":     [{"t","text"}],
+      "instruction": str,
+      "answer":      str | None,
+      "labels":      [T ints],
+    }
+
+Output : the wrapper's raw batch contract:
+    {
+      "series":      [B, max_T, max_C] float tensor (NaN→nan_fill, T pad/trunc),
+      "attributes":  list[str],   # descriptive channel text per sample
+      "timestamps":  list[str],   # range text per sample
+      "events":      list[str],   # event text per sample
+      "question":    list[str],   # from sample["instruction"]
+      "answer":      list[str] | None,
+    }
+
+Variable-length text (token padding / attention_mask) is handled downstream by
+the wrapper's per-sample splice + max_seq padding, so the collator only needs to
+produce the per-sample strings.
+
+Design ref: §6.1 (``data/collator.py``).
+"""
+from __future__ import annotations
+
+from typing import Any, List, Optional
+
+import torch
+
+
+def _attrs_to_text(attrs: List[dict]) -> str:
+    parts = []
+    for a in attrs:
+        parts.append(f"{a.get('name','?')}({a.get('unit','')},{a.get('freq','')})")
+    return " ".join(parts)
+
+
+def _events_to_text(events: List[dict]) -> str:
+    if not events:
+        return ""
+    parts = [f"t={e.get('t')}: {e.get('text','')}" for e in events]
+    return "; ".join(parts)
+
+
+def _timestamps_to_text(ts: List[int]) -> str:
+    if not ts:
+        return ""
+    if len(ts) == 1:
+        return f"t={ts[0]}"
+    return f"t={ts[0]}..{ts[-1]} ({len(ts)} steps)"
+
+
+class Collator:
+    def __init__(self, max_T: int = 512, nan_fill: float = 0.0) -> None:
+        self.max_T = max_T
+        self.nan_fill = nan_fill
+
+    def _stack_series(self, samples: List[dict]) -> torch.Tensor:
+        tensors = []
+        for s in samples:
+            arr = torch.tensor(s["series"], dtype=torch.float32)  # [T, C]
+            T, C = arr.shape
+            # truncate / pad T to max_T
+            if T >= self.max_T:
+                arr = arr[: self.max_T]
+            else:
+                pad = torch.zeros(self.max_T - T, C)
+                arr = torch.cat([arr, pad], dim=0)
+            tensors.append((arr, C))
+        max_C = max(c for _, c in tensors)
+        B = len(tensors)
+        out = torch.zeros(B, self.max_T, max_C, dtype=torch.float32)
+        for i, (arr, C) in enumerate(tensors):
+            out[i, :, :C] = arr[:, :max_C]
+        # replace NaN (missing markers) with fill
+        out = torch.nan_to_num(out, nan=self.nan_fill)
+        return out
+
+    def __call__(self, samples: List[dict]) -> dict:
+        series = self._stack_series(samples)
+        attributes = [_attrs_to_text(s.get("attributes", [])) for s in samples]
+        timestamps = [_timestamps_to_text(s.get("timestamps", [])) for s in samples]
+        events = [_events_to_text(s.get("events", [])) for s in samples]
+        question = [s.get("instruction", "") for s in samples]
+        answers = [s.get("answer") for s in samples]
+        if all(a is None for a in answers):
+            answer: Optional[List[str]] = None
+        else:
+            answer = ["" if a is None else a for a in answers]
+        return {
+            "series": series,
+            "attributes": attributes,
+            "timestamps": timestamps,
+            "events": events,
+            "question": question,
+            "answer": answer,
+        }
diff --git a/tests/test_collator.py b/tests/test_collator.py
new file mode 100644
index 0000000..032b64f
--- /dev/null
+++ b/tests/test_collator.py
@@ -0,0 +1,97 @@
+"""Tests for Collator (T2.5): JSONL sample dicts → wrapper raw batch."""
+import math
+
+import torch
+
+from tsmm.data.collator import Collator
+
+
+def make_sample(T=8, C=3, with_nan=False, n_events=1, answer="no", instruction="Is there an anomaly?"):
+    """Build a dict mimicking the gen_pipeline JSONL schema."""
+    series = [[float(i * C + c) for c in range(C)] for i in range(T)]
+    if with_nan:
+        series[1][0] = float("nan")
+    return {
+        "series": series,                 # [T, C]
+        "attributes": [
+            {"name": f"var{c}", "unit": "%", "freq": "1Hz"} for c in range(C)
+        ],
+        "timestamps": [1700000000 + i for i in range(T)],
+        "events": [{"t": 3, "text": "deploy"}] * n_events,
+        "instruction": instruction,
+        "answer": answer,
+        "labels": [0] * T,
+    }
+
+
+class TestCollator:
+    def test_basic_shapes_and_types(self):
+        col = Collator(max_T=8)
+        batch = col([make_sample(T=8, C=3) for _ in range(4)])
+        assert batch["series"].shape == (4, 8, 3)
+        assert isinstance(batch["attributes"], list) and len(batch["attributes"]) == 4
+        assert isinstance(batch["timestamps"], list) and len(batch["timestamps"]) == 4
+        assert isinstance(batch["events"], list) and len(batch["events"]) == 4
+        assert isinstance(batch["question"], list) and len(batch["question"]) == 4
+        assert isinstance(batch["answer"], list) and len(batch["answer"]) == 4
+
+    def test_question_maps_from_instruction(self):
+        col = Collator()
+        batch = col([make_sample()])
+        assert batch["question"][0] == "Is there an anomaly?"
+
+    def test_answer_passthrough(self):
+        col = Collator()
+        batch = col([make_sample(answer="yes")])
+        assert batch["answer"][0] == "yes"
+
+    def test_no_nan_after_collate(self):
+        col = Collator(nan_fill=0.0)
+        batch = col([make_sample(T=8, C=3, with_nan=True)])
+        assert not torch.isnan(batch["series"]).any()
+        # the NaN position was filled with 0
+        assert batch["series"][0, 1, 0].item() == 0.0
+
+    def test_truncates_long_T(self):
+        col = Collator(max_T=16)
+        batch = col([make_sample(T=32, C=2)])
+        assert batch["series"].shape == (1, 16, 2)
+
+    def test_pads_short_T(self):
+        col = Collator(max_T=64)
+        batch = col([make_sample(T=16, C=2)])
+        assert batch["series"].shape == (1, 64, 2)
+        # padded region is zero
+        assert batch["series"][0, 16:].abs().sum().item() == 0.0
+
+    def test_pads_variable_C_within_batch(self):
+        col = Collator(max_T=16)
+        batch = col([make_sample(T=16, C=2), make_sample(T=16, C=5)])
+        # both padded to max_C=5
+        assert batch["series"].shape == (2, 16, 5)
+
+    def test_attribute_string_is_descriptive(self):
+        col = Collator()
+        batch = col([make_sample(C=2)])
+        s = batch["attributes"][0]
+        assert "var0" in s and "var1" in s
+
+    def test_timestamps_string(self):
+        col = Collator()
+        batch = col([make_sample(T=8)])
+        s = batch["timestamps"][0]
+        assert isinstance(s, str) and "1700000000" in s
+
+    def test_events_string(self):
+        col = Collator()
+        batch = col([make_sample(n_events=2)])
+        s = batch["events"][0]
+        assert isinstance(s, str) and "deploy" in s
+
+    def test_answer_optional(self):
+        # answer=None path (inference): collator should omit/None answer
+        col = Collator()
+        sample = make_sample()
+        sample["answer"] = None
+        batch = col([sample])
+        assert batch.get("answer") is None or batch["answer"][0] is None