diff --git a/README.md b/README.md
index ca9f77e..22e00cd 100644
--- a/README.md
+++ b/README.md
@@ -1,9 +1,20 @@
 # chinese -> english finetuning datasets
 
+## dataset_v3.0_alpaca_noinstr.json
+![sequence distribution](./img/sequence_lengths.png "Sequence distribution")
+
+- 487M
+- Dataset size: 37243 samples
+- Maximum sequence length: 13760
+- Average sequence length: 3123.26
+
+
+## 
+
 train.en and train.zh are from [here](https://www.dropbox.com/scl/fo/dtrf3pe1vfbo5nse16648/ANLqlv3ascANpkdnYF_w4Jk/V1/TRAIN?dl=0&rlkey=486vbn17qra1ez91btj0n4xu2&subfolder_nav_tracking=1)  
 the [actual dataset and .sqlite file](https://mega.nz/folder/byoFHRST#Mcn6-mU5spHxPg0nMlRS3w)
-It's missing the epubs dir I used for paragraph rebuilding... I accidentally deleted the dir, sorry :c  
-What I did was Google a sentence from the chapter 1 of a novel and just scrape 50-60 chapters from either Webnovel or some aggregator, then unzip it into epub with the directory name set to `book_id`.
+
+
 
 GuoFeng dataset chapter spread:
 
diff --git a/img/sequence_lengths.png b/img/sequence_lengths.png
new file mode 100644
index 0000000..cdf4d31
Binary files /dev/null and b/img/sequence_lengths.png differ
diff --git a/sequence_len.py b/sequence_len.py
new file mode 100644
index 0000000..ad4733c
--- /dev/null
+++ b/sequence_len.py
@@ -0,0 +1,59 @@
+from torchtune.data import Message
+from torchtune.models.qwen2 import qwen2_tokenizer
+from prompts.translation import TranslateTemplate
+from tqdm import tqdm
+from tqdm.contrib.concurrent import process_map
+import json
+
+
+def analyze_sequence_lengths(vocab_path, merges_path, json_path):
+    # Load Qwen2 tokenizer
+    tokenizer = qwen2_tokenizer(vocab_path, merges_path)
+    translate_template = TranslateTemplate()
+
+    with open(json_path, "r", encoding="utf-8") as f:
+        dataset = json.load(f)
+
+    max_len = 0
+    lengths = []
+
+    for sample in tqdm(dataset):
+        # Convert sample to messages
+        msgs = [
+            Message(role="user", content=sample["input"]),
+            Message(role="assistant", content=sample["output"]),
+        ]
+
+        templated_msgs = translate_template(msgs)
+
+        # Tokenize messages
+        tokens, mask = tokenizer.tokenize_messages(templated_msgs)
+        seq_len = len(tokens)
+        lengths.append(seq_len)
+        max_len = max(max_len, seq_len)
+
+    avg_len = sum(lengths) / len(lengths)
+    print(f"\nDataset size: {len(dataset)} samples")
+    print(f"Maximum sequence length: {max_len}")
+    print(f"Average sequence length: {avg_len:.2f}")
+
+    # Optional: Plot distribution
+    import matplotlib.pyplot as plt
+
+    plt.figure(figsize=(10, 6))
+    plt.hist(lengths, bins=50)
+    plt.title("Distribution of Sequence Lengths")
+    plt.xlabel("Sequence Length")
+    plt.ylabel("Count")
+    plt.savefig("sequence_lengths.png")  # or .jpg
+    plt.close()
+
+    return max_len, lengths
+
+
+# Example usage
+vocab_path = "/home/mira/models/Qwen2.5-7B-Base/vocab.json"
+merges_path = "/home/mira/models/Qwen2.5-7B-Base/merges.txt"
+dataset = "/home/mira/models/datasets/GuoFeng/datasets/dataset_v3.0_alpaca_noinstr.json"
+
+max_len, lengths = analyze_sequence_lengths(vocab_path, merges_path, dataset)