Update data

README.md

@@ -5,26 +5,24 @@ task_categories:
 language:
 - en
 - zh
-size_categories:
-- n<1K
 configs:
-- config_name: default
+- config_name: chatgpt-2000
   default: true
   data_files:
   - split: train
-    path: "train.json"
-  - split: test
-    path: "test.json"
-- config_name: chatgpt
-  data_files:
-  - split: train
-    path: "chatgpt-train.jsonl"
+    path: "chatgpt-train-2000.jsonl"
   - split: test
     path: "chatgpt-test.jsonl"
-- config_name: chatgpt-recent
+- config_name: chatgpt-8192
   data_files:
   - split: train
-    path: "chatgpt-recent-train.jsonl"
+    path: "chatgpt-train-8192.jsonl"
   - split: test
-    path: "chatgpt-recent-test.jsonl"
----
+    path: "chatgpt-test-8192.jsonl"
+---
+
+This repository holds the data file for translating TechLinked, which talks about mostly technology and science news.
+
+Raw data is in the data/ folder. Scripts generate OpenAI's ChatCompletion Fine-tuning API formatted training data in `jsonl` format.
+
+`-2000` variants are designed to be used with GPT-3 with 8192 tokens context length limit. `-8192` variants are designed to be used with GPT-4o mini with 128000 context window and 16384 max output tokens.

chatgpt-test-2000.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5b2b51eef25ded33a5060bf717dd024285d32d140bebe146f808500a6655e92
+size 142305

chatgpt-test-8192.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07873a323dea01f082dab4e695d24f87dac2471142863c5e2c1167d2593ea2be
+size 133630

chatgpt-train-2000.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b18ee356a6301b7912f142227a389bdcfaf55956814473abb817689f80fb17b9
+size 594567

chatgpt-train-8192.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:832d3d0ed431f4fb73f59e2aab560706553f3813045f5e17dc54b1396f5f88a6
+size 549979

combined-2000.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f11f3da6c70be783b0110e9dd64a28a216937cfc393ecf651e90163be8863f81
+size 736869

combined-8192.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aaddc2f5fb4a2eb2d5c6a9a40bea8ab92f1333ffb0ebc2445bb6ce265ce50e5
+size 683606

data/2023-10-18 China lost its GPU privileges [Wz9JoTTRa4k].cn.txt

@@ -110,8 +110,8 @@
 [292.04] 就算是编辑 他也没法切掉我 [293.8]
 [293.88] 我要继续下去 [294.28]
 [294.36] 纽约的立法者们提出了一项新法案 [296.64]
-[296.64] 要求购买任何可以理论上打印出全套或部分火器的3D打印机 [298.62]
-[298.62] 都需要进行刑事背景调查 [303.55]
+[296.64] 要求购买任何可以理论上打印出全套或部分火器的3D打印机 [298.56]
+[298.62] 都需要进行刑事背景调查 [303.5]
 [303.92] 不巧的是 符合这一条件的打印机占了大多数 [306.26]
 
 [306.42] Gun Digest推荐Creality Ender 3 V2 [310.1]
@@ -126,8 +126,8 @@
 [328.92] AMD发布了一个新的beta显卡驱动程序 [330.82]
 [331.06] 该驱动禁用了Radeon Anti-Lag Plus技术 [333.7]
 
-[333.7] 在该功能及其前身 不带加号的Anti-Lag此前导致了 [338.57]
-[338.58] CS2、Apex Legends 和COD的玩家封禁风波 [343.29]
+[333.7] 在该功能及其前身 不带加号的Anti-Lag此前导致了 [338.58]
+[338.58] CS2、Apex Legends 和COD的玩家封禁风波 [343.3]
 [343.68] 不幸的是 [344.24]
 [344.76] AMD建议使用Anti-Lag技术以减少AFMF造成的延迟 [349.46]
 [349.86] 在将其在所有DX11游戏中启用后 [352.68]
@@ -159,12 +159,11 @@
 [422.58] 当时IRS与税务代理合作 [425.4]
 
 [425.4] 为那些纳税申报相对简单的纳税人 提供了免费的替代方案 [429.94]
-[430.38] 然后这些公司将这些免费替代方案 [432.3]
-[432.62] 变得比找早起理由还要困难 [434.94]
+[430.38] 然后这些公司将这些免费替代方案 变得比找早起理由还要困难 [434.94]
 [435.34] 或者说找晚上上床的理由 [437.04]
 [437.16] 为什么这两个都是？ [437.94]
 [441.0] 当IRS禁止他们这么做时 [442.74]
-[443.02] 万豪和Intuit（TurboTax的制造商） [446.44]
+[443.02] 万豪和Intuit（TurboTax的制造商） [446.08]
 [446.44] 离开了这个协议 [447.16]
 [447.52] Intuit已经游说反对直接申报十多年了 [451.16]
 
@@ -172,9 +171,9 @@
 [453.96] 的确有 但你永远找不到 [456.04]
 [456.32] 那我就继续熬夜了 [458.0]
 [458.16] Slimbook和Fedora项目 [459.46]
-[459.46] 宣布推出了一款针对Linux优化的新款笔记本电脑 Fedora Slimbook [463.32]
-[463.78] 以Fedora Slim为名的 [465.24]
-[465.46] Fedora Slimbook的名字来源 [465.46]
+[459.46] 宣布推出了一款针对Linux优化的新款笔记本电脑 [461.64]
+[462.08] Fedora Slimbook [463.32]
+[463.78] 名字来源于 Fedora Slimbook [465.24]
 [465.46] 是一个铁骨铮铮的私家侦探 [467.02]
 [467.02] 他还是Marlboro的广告牌上的烟枪牛仔 yee-haw! [469.52]
 

data/2023-10-18 China lost its GPU privileges [Wz9JoTTRa4k].en.txt

@@ -47,7 +47,7 @@
 [107.5] without running into walls. [108.92]
 [109.14] But that's half the fun. [110.36]
 [110.52] While that's obviously fine  if you're just taking your $500 Skull accessory  on a walk through the park, [115.42]
-[115.66] it raises obvious privacy concerns [117.56]
+[115.66] it raises obvious privacy concerns [117.55]
 [117.64] when random tech bros are exploring the wonder of AR  by recording banal interactions [122.38]
 [122.38] with identifiable service workers [123.94]
 [123.94] and posting it on the internet. [125.26]
@@ -168,7 +168,7 @@
 
 [425.4] to offer taxpayers with simple returns [427.72]
 [427.72] a free alternative for filing their taxes. [429.94]
-[430.38] Those companies then made those free alternatives  harder to find than a reason to get up in the morning [434.94]
+[430.38] Those companies then made those free alternatives harder to find than a reason to get up in the morning [434.94]
 [435.34] or a reason to go to bed at night. [437.04]
 [437.16] Why is it both? [437.94]
 [441.0] When the IRS banned them from doing that, [442.74]

data/2023-10-25 Windows' Apple Silicon Moment is Coming [xFvGERpRUpM].cn.txt

@@ -40,8 +40,8 @@
 
 [129.4] 多位消息人士称NVIDIA正在准备一款RTX 4070 Super [132.78]
 [132.78] 和一款拥有20GB显存的RTX 4080的Super版 [137.38]
-[137.82] NVIDIA自从2020年4月发布RTX 2080 Super以来 [140.87]
-[140.87] 就再也没有使用过"Super"这个名号 [143.8]
+[137.82] NVIDIA自从2020年4月发布RTX 2080 Super以来 [140.16]
+[140.16] 就再也没有使用过"Super"这个名号 [143.8]
 [144.3] 也许是因为NVIDIA进入了一个更加 [146.82]
 [147.42] 坏人的阶段 [148.34]
 [148.76] 所以不值得Super的名号 [149.52]

data/2023-10-28 Google Pays Apple EVERY YEAR [qvZAYJLmzuk].cn.txt

@@ -66,7 +66,7 @@
 [185.48] 现在可能对他们将不得不重做的文书工作感到恼火 [188.44]
 [188.44] 因为他们批准了127亿美元 [191.66]
 [191.66] 为合并提供资金援助 [193.54]
-[195.74] 哦 [196.6]
+[195.74] 哦 [196.36]
 [196.6] 爱的魔力转圈圈 [198.3]
 
 [198.56] 这不是第一次西数试图与铠侠合并 [202.84]

generate_chatgpt_varlen.py

@@ -0,0 +1,214 @@
+import json
+import os
+import copy
+import re
+from os import listdir
+from os.path import isfile, join
+import argparse
+import sys
+
+INSTRUCT_CHUNKED_PROMPT = """你是一个擅长翻译科技新闻的翻译专家。请将以下内容翻译为中文，使用相同格式输出，并保留时间戳。不要漏掉任何信息。合并多行文本时，保留第一个和最后一个时间戳。
+"""
+
+def break_line(line: str):
+    pattern = re.compile(r"^\[(\d+.\d+)\](.*)\[(\d+.\d+)\]$")
+    match = pattern.match(line)
+    start_time = match.group(1)
+    text = match.group(2)
+    end_time = match.group(3)
+    return start_time, text.strip(), end_time, float(start_time), float(end_time)
+
+def get_total_chars(cn_lines: list[str], en_lines: list[str]):
+    cn_total_chars = 0
+    en_total_chars = 0
+
+    for line in cn_lines:
+        cn_total_chars += len(line)
+    
+    for line in en_lines:
+        en_total_chars += len(line)
+    
+    return cn_total_chars + en_total_chars
+
+def chunk_messages(cn_lines: list[str], en_lines: list[str], MAX_LEN: int = 2000):
+    cn_lines_copy = copy.deepcopy(cn_lines)
+    en_lines_copy = copy.deepcopy(en_lines)
+
+    final_chunks: list[tuple[list[str], list[str]]] = []
+    
+    while True:
+        en_current_chunk = []
+        cn_current_chunk = []
+        while True:
+            curr_total_len = get_total_chars(cn_current_chunk, en_current_chunk)
+            if len(cn_lines_copy) == 0 or len(en_lines_copy) == 0:
+                final_chunks.append((cn_current_chunk, en_current_chunk))
+                return final_chunks
+            elif curr_total_len > MAX_LEN:
+                final_chunks.append((cn_current_chunk, en_current_chunk))
+                break
+            else:
+                # Try append a new line to current chunk
+                latest_cn_line = cn_lines_copy.pop(0)
+                cn_start, cn_text, cn_end, cn_start_f, cn_end_f = break_line(latest_cn_line)
+                cn_current_chunk.append(latest_cn_line)
+                while True:
+                    latest_en_line = en_lines_copy.pop(0)
+                    en_start, en_text, en_end, en_start_f, en_end_f = break_line(latest_en_line)
+                    en_current_chunk.append(latest_en_line)
+                    if en_end == cn_end:
+                        break
+                    else:
+                        if en_start_f > cn_end_f:
+                            raise Exception("English and Chinese lines are not in sync. Offensing line: " + latest_cn_line)
+                
+def new_message(eng_in, chs_out, prev_in = None, prev_out = None):
+    if(prev_in == None or prev_out == None):
+        return {"messages": [
+            {"role": "system", "content": INSTRUCT_CHUNKED_PROMPT}, 
+            {"role": "user", "content": eng_in}, 
+            {"role": "assistant", "content": chs_out}]
+        }
+    else:
+        return {"messages": [
+            {"role": "system", "content": INSTRUCT_CHUNKED_PROMPT}, 
+            {"role": "user", "content": prev_in}, 
+            {"role": "assistant", "content": prev_out},
+            {"role": "user", "content": eng_in}, 
+            {"role": "assistant", "content": chs_out}]
+        }
+
+def write_jsonl(message_groups, filename):
+    json_lines = []
+    with open(filename, "w", encoding='utf-8-sig') as fout:
+        for i in range(len(message_groups)):
+            if(i>0):
+                msg_obj = new_message(
+                    message_groups[i][0].strip(),
+                    message_groups[i][1].strip(),
+                    message_groups[i-1][0].strip(),
+                    message_groups[i-1][1].strip()
+                )
+            else:
+                msg_obj = new_message(
+                    message_groups[i][0].strip(),
+                    message_groups[i][1].strip()
+                )
+            json.dump(msg_obj, fout)
+            fout.write("\n")
+            json_lines.append(json.dumps(msg_obj))
+    return json_lines
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Generate ChatGPT training data from a directory of subtitle files.')
+    parser.add_argument('data_dir', type=str, help='The directory containing the subtitle files.', default="data")
+    parser.add_argument('--maxlen', type=int, help='The maximum length of a combined message. \nNote that this limit will be exceeded a little bit, so leave some headroom. \nRecommended value is max context length / 4.', default=2000)
+    parser.add_argument('--test-ratio', type=float, help='The ratio of test data to training data.', default=0.2)
+
+    args = parser.parse_args()
+
+    message_groups = []
+
+    DOCUMENT_ROOT = args.data_dir
+    files = listdir(DOCUMENT_ROOT)
+    files = list(filter(lambda x: x.endswith(".en.txt"), files))
+    files.sort()
+
+    print(files)
+
+    for f in files:
+        en_fname = join(DOCUMENT_ROOT, f)
+        if en_fname.endswith(".en.txt") and isfile(en_fname):
+            cn_fname = join(DOCUMENT_ROOT, f.replace(".en.txt", ".cn.txt"))
+            if os.path.exists(cn_fname) and isfile(cn_fname):
+                print(f"Found data pair: {en_fname} and {cn_fname}")
+
+                with open(en_fname, "r", encoding='utf-8-sig') as enfin:
+                    en_messages = enfin.read()
+
+                with open(cn_fname, "r", encoding='utf-8-sig') as cnfin:
+                    cn_messages = cnfin.read()
+                
+                en_messages = [part.strip() for part in en_messages.split("\n") if part.strip() != ""]
+                cn_messages = [part.strip() for part in cn_messages.split("\n") if part.strip() != ""]
+
+                try:
+                    chunks = chunk_messages(cn_messages, en_messages, MAX_LEN=args.maxlen)
+                    en_messages = []
+                    cn_messages = []
+                    for chunk in chunks:
+                        cn_chunk, en_chunk = chunk
+                        en_messages.append("\n".join(en_chunk))
+                        cn_messages.append("\n".join(cn_chunk))
+                        print("\n".join(en_chunk))
+                        print("---")
+                        print("\n".join(cn_chunk))
+                        print("\n")
+                except Exception as e:
+                    print(f"Error: {e}")
+                    continue
+                
+                if(len(en_messages) != len(cn_messages)):
+                    print(f"English and Chinese version mismatch. Discarding {en_fname} pair.")
+                
+                messages = zip(en_messages, cn_messages)
+
+                message_groups.extend(messages)
+
+    jsonl_lines = write_jsonl(message_groups, f"combined-{args.maxlen}.jsonl")
+
+    import random
+    random.shuffle(jsonl_lines)
+
+    TEST_RATIO = args.test_ratio
+
+    split_index = int(len(jsonl_lines) * TEST_RATIO)
+
+    test = jsonl_lines[:split_index]
+    train = jsonl_lines[split_index:]
+
+    with open (f"chatgpt-train-{args.maxlen}.jsonl", "w", encoding='utf-8-sig') as fout:
+        for line in train:
+            fout.write(line + "\n")
+
+    with open (f"chatgpt-test-{args.maxlen}.jsonl", "w", encoding='utf-8-sig') as fout:
+        for line in test:
+            fout.write(line + "\n")
+
+    # recent_files = files[-5:]
+    # recent_messages = []
+
+    # for f in recent_files:
+    #     en_fname = join(DOCUMENT_ROOT, f)
+    #     if en_fname.endswith(".en.txt") and isfile(en_fname):
+    #         cn_fname = join(DOCUMENT_ROOT, f.replace(".en.txt", ".cn.txt"))
+    #         if os.path.exists(cn_fname) and isfile(cn_fname):
+    #             print(f"Found data pair: {en_fname} and {cn_fname}")
+
+    #             with open(en_fname, "r", encoding='utf-8-sig') as enfin:
+    #                 en_messages = enfin.read()
+
+    #             with open(cn_fname, "r", encoding='utf-8-sig') as cnfin:
+    #                 cn_messages = cnfin.read()
+                
+    #             en_messages = [part.strip() for part in en_messages.split("\n") if part.strip() != ""]
+    #             cn_messages = [part.strip() for part in cn_messages.split("\n") if part.strip() != ""] 
+
+    #             if(len(en_messages) != len(cn_messages)):
+    #                 print(f"English and Chinese version mismatch. Discarding {en_fname} pair.")
+                
+    #             messages = zip(en_messages, cn_messages)
+
+    #             recent_messages.extend(messages)
+
+    # write_jsonl(recent_messages, "recent-combined.jsonl")
+
+    # TEST_RATIO = 0.2
+
+    # split_index = int(len(recent_messages) * TEST_RATIO)
+
+    # test = recent_messages[:split_index]
+    # train = recent_messages[split_index:]
+
+    # write_jsonl(train, "chatgpt-recent-train.jsonl")
+    # write_jsonl(test, "chatgpt-recent-test.jsonl")