Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+chat_sprint_artifacts/corpus.txt filter=lfs diff=lfs merge=lfs -text

chat.py

@@ -0,0 +1,288 @@
+#!/usr/bin/env python
+# chat.py — Gradio chat app with **turn-taking guards**
+# ----------------------------------------------------
+# pip install torch sentencepiece gradio
+# python chat.py
+#
+# Expects in ./chat_sprint_artifacts (or CHAT_SPRINT_ARTIFACTS env):
+#   - spm_chat.model
+#   - tinygpt.pt
+#   - model_config.json
+#
+# Fixes for your screenshot:
+#   • Masks “You:” and tag tokens so the bot can’t emit them mid-reply.
+#   • Stops generation on EOS / "\n\n" / "\nYou:" so it doesn’t start the next turn.
+#   • Works with Gradio 5.x (type="messages", queue() w/o kwargs).
+
+import os, json, math
+from pathlib import Path
+from typing import List, Tuple, Generator
+
+import torch
+import torch.nn as nn
+import sentencepiece as spm
+import gradio as gr
+
+# ---------- paths & device ----------
+ART = Path(os.environ.get("CHAT_SPRINT_ARTIFACTS", "chat_sprint_artifacts"))
+SPM_PATH = ART / "spm_chat.model"
+CKPT     = ART / "tinygpt.pt"
+CFG_JSON = ART / "model_config.json"
+LOG_PATH = ART / "chat_transcript.txt"
+ART.mkdir(parents=True, exist_ok=True)
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+if DEVICE.type == "cuda":
+    torch.set_float32_matmul_precision("high")
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+# ---------- tiny GPT (same as trainer) ----------
+class CausalSelfAttention(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        assert n_embd % n_head == 0
+        self.n_head = n_head
+        self.head_dim = n_embd // n_head
+        self.qkv = nn.Linear(n_embd, 3*n_embd, bias=False)
+        self.proj = nn.Linear(n_embd, n_embd, bias=False)
+        self.attn_drop = nn.Dropout(dropout)
+        self.resid_drop = nn.Dropout(dropout)
+        self.register_buffer("mask", torch.tril(torch.ones(block_size, block_size)).view(1,1,block_size,block_size))
+    def forward(self, x):
+        B,T,C = x.shape
+        qkv = self.qkv(x); q,k,v = qkv.chunk(3, dim=-1)
+        q = q.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        k = k.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        v = v.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        att = (q @ k.transpose(-2,-1)) / math.sqrt(self.head_dim)
+        att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
+        att = torch.softmax(att, dim=-1)
+        att = self.attn_drop(att)
+        y = att @ v
+        y = y.transpose(1,2).contiguous().view(B,T,C)
+        y = self.resid_drop(self.proj(y))
+        return y
+
+class Block(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(n_embd)
+        self.attn = CausalSelfAttention(n_embd, n_head, dropout, block_size)
+        self.ln2 = nn.LayerNorm(n_embd)
+        self.mlp = nn.Sequential(
+            nn.Linear(n_embd, 4*n_embd),
+            nn.GELU(),
+            nn.Linear(4*n_embd, n_embd),
+            nn.Dropout(dropout),
+        )
+    def forward(self, x):
+        x = x + self.attn(self.ln1(x))
+        x = x + self.mlp(self.ln2(x))
+        return x
+
+class TinyGPT(nn.Module):
+    def __init__(self, vocab_size, n_layer, n_head, n_embd, block_size, dropout=0.0):
+        super().__init__()
+        self.block_size = block_size
+        self.tok_emb = nn.Embedding(vocab_size, n_embd)
+        self.pos_emb = nn.Embedding(block_size, n_embd)
+        self.blocks = nn.ModuleList([Block(n_embd, n_head, dropout, block_size) for _ in range(n_layer)])
+        self.ln_f = nn.LayerNorm(n_embd)
+        self.head = nn.Linear(n_embd, vocab_size, bias=False)
+        self.apply(self._init)
+    def _init(self, m):
+        if isinstance(m, (nn.Linear, nn.Embedding)): nn.init.normal_(m.weight, mean=0.0, std=0.02)
+        if isinstance(m, nn.Linear) and m.bias is not None: nn.init.zeros_(m.bias)
+        if isinstance(m, nn.LayerNorm): nn.init.ones_(m.weight); nn.init.zeros_(m.bias)
+    def forward(self, idx):
+        B,T = idx.shape; assert T <= self.block_size
+        pos = torch.arange(0, T, device=idx.device)
+        x = self.tok_emb(idx) + self.pos_emb(pos)[None,:,:]
+        for blk in self.blocks: x = blk(x)
+        x = self.ln_f(x)
+        logits = self.head(x)
+        return logits
+
+    @torch.no_grad()
+    def sample_stream(
+        self, idx, sp: spm.SentencePieceProcessor,
+        forbid_ids=None, stop_ids=None, stop_strings=None,
+        max_new_tokens=200, temperature=0.8, top_k=60, top_p=0.95, repetition_penalty=1.0
+    ) -> Generator[str, None, None]:
+        """Yield decoded text chunks with simple constraints."""
+        forbid_ids = set(forbid_ids or [])
+        stop_ids   = set(stop_ids or [])
+        stop_strings = list(stop_strings or [])
+        prev_text = sp.decode(idx[0].tolist())
+
+        for _ in range(int(max_new_tokens)):
+            idx_cond = idx[:, -self.block_size:]
+            logits = self.forward(idx_cond)[:, -1, :]
+
+            # role/tag masking
+            if forbid_ids:
+                mask_idx = torch.tensor(list(forbid_ids), device=logits.device)
+                logits[:, mask_idx] = -1e9
+
+            # repetition penalty
+            if repetition_penalty != 1.0:
+                uniq, _ = torch.unique(idx_cond[0], return_counts=True)
+                logits[:, uniq] /= repetition_penalty
+
+            # temp / top-k / top-p
+            logits = logits / max(1e-8, float(temperature))
+            if top_k and int(top_k) > 0:
+                v, _ = torch.topk(logits, min(int(top_k), logits.size(-1)))
+                cutoff = v[:, -1].unsqueeze(-1)
+                logits = torch.where(logits < cutoff, torch.full_like(logits, -1e9), logits)
+            if top_p and float(top_p) < 0.9999:
+                sorted_logits, sorted_idx = torch.sort(logits, descending=True)
+                probs = torch.softmax(sorted_logits, dim=-1)
+                cdf = torch.cumsum(probs, dim=-1)
+                mask = cdf > float(top_p); mask[:, 0] = False
+                sorted_logits[mask] = -1e9
+                logits = torch.zeros_like(logits).scatter(1, sorted_idx, sorted_logits)
+
+            # sample one
+            probs = torch.softmax(logits, dim=-1)
+            next_id = torch.multinomial(probs, num_samples=1)
+
+            # token-level stops
+            if int(next_id) in stop_ids:
+                break
+
+            idx = torch.cat([idx, next_id], dim=1)
+
+            # text-level stops
+            full_text = sp.decode(idx[0].tolist())
+            new_chunk = full_text[len(prev_text):]
+            acc_text = full_text  # for string-stop check
+            cut_at = None
+            for s in stop_strings:
+                pos = acc_text.find(s, len(prev_text))  # only search in newly produced tail
+                if pos != -1:
+                    cut_at = pos
+                    break
+            if cut_at is not None:
+                # yield only up to stop sequence, then stop
+                yield acc_text[len(prev_text):cut_at-len(prev_text)]
+                break
+
+            yield new_chunk
+            prev_text = full_text
+
+# ---------- artifacts ----------
+def load_everything():
+    if not SPM_PATH.exists(): raise FileNotFoundError(f"Missing tokenizer: {SPM_PATH}")
+    if not CKPT.exists():     raise FileNotFoundError(f"Missing weights: {CKPT}")
+    if not CFG_JSON.exists(): raise FileNotFoundError(f"Missing config: {CFG_JSON}")
+
+    sp = spm.SentencePieceProcessor(); sp.load(str(SPM_PATH))
+    cfg = json.loads(CFG_JSON.read_text())
+    model = TinyGPT(
+        vocab_size=cfg["vocab_size"],
+        n_layer=cfg["n_layer"], n_head=cfg["n_head"], n_embd=cfg["n_embd"],
+        block_size=cfg["block_size"], dropout=cfg.get("dropout", 0.0)
+    ).to(DEVICE)
+    sd = torch.load(CKPT, map_location=DEVICE)
+    model.load_state_dict(sd, strict=True)
+    model.eval()
+    return sp, model, cfg
+
+SP, MODEL, CFG = load_everything()
+
+# special pieces we want to control at decode-time
+EOS_ID   = SP.eos_id()
+YOU_ID   = SP.piece_to_id("You:")
+BOT_ID   = SP.piece_to_id("Bot:")
+TAGS     = ["[STYLE=Snark]", "[FORM=TWEET]", "[FORM=HEADLINE]", "[MOOD=Unhinged]", "[MOOD=Cheeky]"]
+TAG_IDS  = [SP.piece_to_id(t) for t in TAGS if SP.piece_to_id(t) != -1]
+FORBID_IDS = {x for x in [YOU_ID, BOT_ID] + TAG_IDS if x != -1}  # forbid these inside assistant text
+STOP_IDS   = {i for i in [EOS_ID, YOU_ID] if i != -1}
+STOP_STRS  = ["\nYou:", "\n\n"]   # treat next turn / blank-line as stop
+
+# ---------- prompt building ----------
+STYLE_TAGS = ["", *TAGS]
+
+def history_to_pairs(history_messages) -> List[Tuple[str,str]]:
+    pairs: List[Tuple[str,str]] = []
+    last_user = None
+    for m in history_messages:
+        role = m.get("role"); content = m.get("content", "")
+        if role == "user": last_user = content
+        elif role == "assistant" and last_user is not None:
+            pairs.append((last_user, content)); last_user = None
+    return pairs
+
+def build_prompt(history_pairs: List[Tuple[str,str]], user: str, style: str) -> str:
+    lines = []
+    for u, b in history_pairs:
+        lines.append(f"You: {u}")
+        lines.append(f"Bot: {b}")
+        lines.append("")
+    lines.append(f"You: {user}")
+    if style: lines.append(style)  # we *insert* tags here, but we forbid them inside the reply
+    lines.append("Bot:")
+    return "\n".join(lines)
+
+def encode_ctx(text: str, block_size: int) -> torch.Tensor:
+    ids = SP.encode(text, out_type=int)
+    return torch.tensor([ids[-block_size:]], dtype=torch.long, device=DEVICE)
+
+# ---------- gradio handler ----------
+def respond(message, history, temperature, top_k, top_p, repetition_penalty, max_new_tokens, style):
+    if isinstance(message, dict):
+        message = message.get("content", "")
+    pairs = history_to_pairs(history)
+    prompt = build_prompt(pairs, message, style)
+    x = encode_ctx(prompt, CFG["block_size"])
+
+    stream = MODEL.sample_stream(
+        x, SP,
+        forbid_ids=FORBID_IDS,
+        stop_ids=STOP_IDS,
+        stop_strings=STOP_STRS,
+        max_new_tokens=int(max_new_tokens),
+        temperature=float(temperature),
+        top_k=int(top_k),
+        top_p=float(top_p),
+        repetition_penalty=float(repetition_penalty),
+    )
+
+    acc = ""
+    for chunk in stream:
+        acc += chunk
+        yield acc
+
+    try:
+        with LOG_PATH.open("a", encoding="utf-8") as f:
+            f.write(f"You: {message}\n")
+            if style: f.write(style + "\n")
+            f.write(f"Bot: {acc}\n\n")
+    except Exception:
+        pass
+
+# ---------- app ----------
+def main():
+    title = "TinyGPT — Fun Chat (turn-taking fixed)"
+    desc = f"Device: {DEVICE.type.upper()} • vocab={CFG['vocab_size']} • layers={CFG['n_layer']} heads={CFG['n_head']} dim={CFG['n_embd']} • block={CFG['block_size']}"
+
+    iface = gr.ChatInterface(
+        fn=respond,
+        title=title,
+        description=desc,
+        additional_inputs=[
+            gr.Slider(0.2, 1.5, value=0.8,  step=0.05, label="Temperature"),
+            gr.Slider(0,   200, value=60,   step=1,    label="Top-K (0=off ⇒ set 0)"),
+            gr.Slider(0.5, 1.0, value=0.95, step=0.01, label="Top-P"),
+            gr.Slider(1.0, 1.5, value=1.08, step=0.01, label="Repetition penalty"),
+            gr.Slider(16,  512, value=200,  step=8,    label="Max new tokens"),
+            gr.Dropdown(STYLE_TAGS, value="", label="Style tag"),
+        ],
+        type="messages",
+    )
+    iface.queue()
+    iface.launch(server_name="0.0.0.0", server_port=7860, show_api=False, inbrowser=False)
+
+if __name__ == "__main__":
+    main()

chat_sprint_artifacts/chat_transcript.txt

@@ -0,0 +1,33 @@
+You: tell me a joke
+Bot:  None of the old people are making jokes about how there is this when you get it.
+
+You: how old are you?
+Bot:  I'm 32, thanks for asking.
+
+You: when is your birthday?
+Bot:  It'll be 20 degrees.
+
+You: on your birthday?
+Bot:  It's only a single one with everything that's like.
+
+You: lots of days are 20 degrees
+Bot:  Here are all 10-15 degrees below zero less than a year or worst: 1.
+
+You: sounds cold
+Bot:  Can you tell me more about why you’re so cold?
+
+You: talk to me like shakespeare
+Bot:  Why do you think they are smelly and disgusting?
+
+You: no, i like him
+Bot:  Do you mean “we know my names” instead of “All other people”?
+
+You: what?
+Bot:  I was just looking at you.
+
+You: where is the first astronaut buried
+Bot:  100.
+
+You: 100 what?
+Bot:  Chris James Between Star Wars a Drive Drive!
+

chat_sprint_artifacts/corpus.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ce34bd85535c80a48cc6f6efe53454f8d1ac54eb6f39e53c25c7666bc8e5d4c
+size 20843161

chat_sprint_artifacts/model_config.json

@@ -0,0 +1,8 @@
+{
+  "vocab_size": 5000,
+  "n_layer": 6,
+  "n_head": 6,
+  "n_embd": 384,
+  "block_size": 256,
+  "dropout": 0.0
+}

chat_sprint_artifacts/samples.txt

@@ -0,0 +1,20 @@
+----
+You: Give me a spicy take on AI. Bot: What should I do to prevent asthma? You: Continue in Shakespearean style. Bot: Or so, I had been borne it; and You: Reply with sarcasm: [STYLE=Snark] Bot: Reddit Comment: The FBIs and Mona comments are bullshit, but the only reason we buy them at a patria are. You: It is a long story, but could you please give me some tips on how to write an essay on the impact of social media on teenagers? Bot: Sure, here are some tips on how to write an essay on social media sharing and retaining your podcast on coworkers on social You: ['i think i have ever heard i got a 3ds', 'well that lot is better off what you like to do'] Bot: My dad and my grandma loves soccer. What kind of Razer would he like?
+----
+You: Roast my messy desk. Bot: What does that mean? Is that your friend you've got a job? You: Yeah, I know. It's been pretty busy at work lately. But I'm trying to stay positive and keep this up. Bot: That's true. So what have you been up to lately? You: Wow, I can't believe I just burned into the group myself. It was so unexpected. Bot: Yeah, it was pretty amazing. And now, I'm really glad we did this together. It feels like just yesterday we were running away. You: Well, you'll never know if you tell me or not, but if you say so. Bot: Okay, then I support you. Thank you for telling me that. You: What did the two blind people say to the other? Bot: A black man walks out of his ear. You: Hey
+----
+You: Explain recursion like you're annoyed. Bot: Thank you for your kind words, I really appreciate it. You: What do you mean? Bot: Well, I just feel like there's more to life than what we have here. You: Okay, well let's take a look at your test and see where we can cut back if we find it. Bot: Yeah, that would be a good idea. Thank you so much for your help! You: Reply with sarcasm: [STYLE=Snark] Bot: Reddit Comment: You don't need any of them. You: Continue in Shakespearean style. Bot: Into the king and heralds are full of tears. You: Reply with sarcasm: [STYLE=Snark] Bot: User: I got ten points on the way back and I can pick up everything from now. You: Yes , I am very grateful . Bot: What time
+----
+You: Write a satirical headline about coffee. Bot: It's an old language, a branch of computer screen that has been deposed to many people, using a larger, a You: That's definitely the right attitude for me. But you're not alone in this. Bot: Yeah, I guess so. But it's just so hard sometimes. You: No problem, man. So, what have you been up to lately? Bot: Just hanging out with friends and family, like I said. We went to the park yesterday and we had a lot of fun. You: Continue a dad joke: Bot: I said “This is at the gate. You: 800 people are going to buy a parasol for any amount of stuff. Can you help me join a trade barbecue?????????????? [STYLE=Snark] Bot: Lol but it's really 34 
+----
+You: Give me a shower thought about umbrellas. Bot: I'm sure some of them will come out of their heads and see if the ads are vaginated from predators You: Thank you so much for your help, boss. I really appreciate it. Bot: You're welcome, Haylea. I'm happy to have you on your team. You: Yeah, I was just thinking the same thing. We were lucky to be alive and taken each other. Bot: Well, I know that the average part was unfair, but I'm not entirely sure. What do you think? You: I found it really fascinating how these art and architecture have shaped the food I gained from the early 1990s. Bot: Yes, the environmental impact of the new and cultural significance of the new season tourism industry is: 1. You: Give me a shower thought. Bot: Do gay people get paid when you
+----
+You: Tell me a one-liner about deadlines. Bot: I gotta make my own reverse You: What do you call a guy who has eve? Bot: Depending. You: How did the Renaissance view of the Age of Place have impacted the indigenous layers of tropicality, and what were some of their con Bot: The Age of Place of Place had a significant impact on the indigenous communities. You: Give me a shower thought. Bot: The radio has shirts and the internet ever already has the same signs/all to do anything with them You: That's really admirable. I can tell that you have very well. Bot: Yeah, I definitely have. So, how long have you been playing sports? You: Continue in Shakespearean style. Bot: And in this thund
+----
+You: Stay in Shakespeare mode and flatter me. Bot: You want to do more than try to use a fake, in a way that looks at the deaf, but you need to be careful with the You: I'm not sure yet. I might go visit my family or a do it. What about you? Bot: Oh, I don't really have anything that I can do. You: A man standing up for a resource Bot: He picked on the jawline. He picked up the man and says "What'll I look like? You: Can you please add some suggestions for color combinations that would work well for each type of exercise as well? Bot: Sure, here are some color schemes that would work well for each type of exercise as well: 1. You: I don't know what I would have done if I had just. Bot: Well, you did. And now I'm
+----
+You: Reply sarcastically to: I love meetings. Bot: So, what are the other things you're interested in doing with your paintings? You: Why did the teacher choose a marriage of the Mumbai? Bot: They can't be between the two people. You: Continue a dad joke: Bot: He buried me that I was a white male, I would say 'ghould'? You: I've been good. Just busy with work and stuff. How about you? Bot: Same here. It's great to see you. How about you? You: I wanted to return this email address for my company. Bot: Sure, how old is he? You: Why do you want to rob a bank? Bot: I’m afraid I can’t do that. I’m not sure who would want to do that. You: Well, I've always wanted to
+----
+You: What's a good way to say no to a meeting politely? Bot: It is best for the children to stay on top of the rest of us. You: I was thinking of straightening it, but I think it would be a good idea if I just take it. Bot: Well, let me check with our other two pieces of furniture and see if anything catches your eye. You: ['got a meme show . i had to go to a tattoo .', "well you have to buy a new full cookie in school ."] Bot: I've never tried that. How about with a large amount of food and water will it affect? You: Hey, did you hear that noise? Bot: Yeah, it's pretty cold out tonight. You: What if you're not allowed on getting married or not? Bot: That’s not an excuse, I’m afraid I can’t give you advice
+----
+You: Roleplay as my productivity coach for two turns. Bot: Optimum, I appreciate your kind words. You: Continue a dad joke: Bot: But I was a jar! You: A young man walks into a bar Bot: Son's rather short and says how the bartender looks up and says "I have some good toilet paper?" Son: "Awwwwww! You: Continue in Shakespearean style. Bot: And tails all sorts of shape, You: Can you please provide more information about how I can integrate the Twitter marketing plan? Bot: Sure, I'd be happy to help you with that. Here are some ways you can promote the Twitter marketing plan: 1. You: I know, but it's just so frustrating. I feel like I'm never going to win or make a mistake. Bot: Maybe you're not being bored, Daryl. But are

chat_sprint_artifacts/spm_chat.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b83805b03803e8c05b162483d36eb016bfb285fd804cbe2878829ceb3c69b572
+size 317857

chat_sprint_artifacts/spm_chat.vocab

@@ -0,0 +1,5000 @@
+<unk>	0
+<s>	0
+</s>	0
+You:	0
+Bot:	0
+[STYLE=Snark]	0
+[FORM=TWEET]	0
+[FORM=HEADLINE]	0
+[MOOD=Unhinged]	0
+[MOOD=Cheeky]	0
+▁	-1.96775
+.	-3.1534
+s	-3.51788
+▁I	-3.69735
+,	-3.71605
+'	-3.73051
+▁a	-4.03037
+▁to	-4.1268
+▁you	-4.14381
+▁the	-4.14513
+?	-4.47509
+t	-4.55324
+▁and	-4.69105
+▁in	-4.78519
+▁of	-4.79873
+▁it	-4.82506
+m	-4.88252
+▁that	-4.96709
+:	-5.09343
+!	-5.15476
+▁for	-5.15792
+▁is	-5.17387
+▁with	-5.21771
+▁me	-5.35357
+re	-5.40692
+ing	-5.54584
+▁can	-5.56813
+▁be	-5.57056
+an	-5.5715
+▁are	-5.57301
+▁have	-5.58497
+▁do	-5.59701
+▁on	-5.62945
+▁my	-5.65631
+▁Continue	-5.68236
+-	-5.70966
+▁like	-5.74127
+d	-5.74409
+▁just	-5.74907
+▁about	-5.76572
+ed	-5.78997
+▁so	-5.83787
+▁know	-5.85083
+▁was	-5.85437
+▁your	-5.88187
+▁not	-5.9076
+▁It	-5.91462
+▁What	-5.99908
+▁what	-6.00905
+▁really	-6.01693
+’	-6.02848
+▁That	-6.04855
+▁don	-6.06456
+▁but	-6.08073
+▁this	-6.08915
+ve	-6.128
+▁You	-6.14882
+▁we	-6.1491
+▁style	-6.15094
+▁Shakespeare	-6.18507
+▁how	-6.21798
+▁some	-6.23558
+ll	-6.23823
+▁one	-6.26434
+y	-6.29163
+▁been	-6.29645
+▁Yeah	-6.33878
+▁think	-6.33896
+▁out	-6.34988
+▁want	-6.40554
+▁all	-6.43015
+▁more	-6.43149
+▁good	-6.43445
+▁joke	-6.47366
+er	-6.47508
+▁The	-6.48227
+▁get	-6.4855
+e	-6.4944
+▁if	-6.50167
+▁they	-6.51369
+▁up	-6.51392
+▁at	-6.51601
+▁dad	-6.53622
+▁great	-6.54628
+▁would	-6.54781
+▁Well	-6.58407
+▁sarcasm	-6.58466
+▁Reply	-6.58792
+▁there	-6.59345
+ly	-6.60068
+▁A	-6.61641
+▁Can	-6.62726
+▁from	-6.64539
+▁But	-6.65102
+▁or	-6.66291
+n	-6.68929
+▁time	-6.70449
+▁here	-6.72834
+▁going	-6.73623
+▁How	-6.76174
+i	-6.76857
+▁too	-6.76894
+▁people	-6.77556
+▁he	-6.77938
+r	-6.78092
+▁see	-6.78676
+▁make	-6.80848
+in	-6.81138
+▁help	-6.82631
+▁Yes	-6.8368
+▁could	-6.85967
+I	-6.85985
+▁thought	-6.86412
+▁much	-6.87212
+▁sure	-6.87562
+al	-6.8784
+▁need	-6.88587
+▁lot	-6.89281
+▁"	-6.90667
+▁go	-6.91176
+▁So	-6.91539
+▁love	-6.92668
+▁feel	-6.93411
+▁has	-6.9642
+▁when	-6.96629
+▁will	-6.97469
+▁right	-6.98445
+▁now	-6.99274
+▁work	-6.99997
+...	-7.00481
+p	-7.01682
+▁by	-7.02233
+▁as	-7.03334
+a	-7.05217
+▁who	-7.06742
+▁them	-7.06782
+▁No	-7.07072
+▁any	-7.07339
+▁did	-7.07361
+▁We	-7.07671
+▁no	-7.07744
+▁sounds	-7.07872
+o	-7.08359
+▁Do	-7.08657
+▁had	-7.09694
+le	-7.10885
+▁always	-7.11434
+▁an	-7.11867
+▁their	-7.15531
+▁something	-7.16866
+c	-7.17042
+es	-7.18271
+▁way	-7.18766
+▁shower	-7.19119
+or	-7.20171
+w	-7.20368
+▁were	-7.20726
+▁her	-7.21555
+▁And	-7.22995
+▁Reddit	-7.23588
+▁Give	-7.2435
+"	-7.25579
+▁sharp	-7.25599
+▁new	-7.2628
+l	-7.2632
+▁his	-7.26459
+▁other	-7.26489
+▁doing	-7.26587
+▁#	-7.26653
+▁Sure	-7.26697
+▁React	-7.27131
+▁User	-7.27181
+liner	-7.27202
+▁i	-7.27481
+▁Comment	-7.28027
+it	-7.28097
+▁should	-7.29053
+▁take	-7.29416
+▁Oh	-7.31442
+▁Thanks	-7.31484
+▁Just	-7.34903
+th	-7.36579
+▁Why	-7.36913
+▁tell	-7.37
+▁well	-7.37299
+▁am	-7.37425
+en	-7.37508
+▁Hey	-7.38031
+▁say	-7.38033
+▁['	-7.38034
+▁sorry	-7.38758
+g	-7.40627
+S	-7.41159
+▁mean	-7.41256
+▁our	-7.42644
+▁This	-7.43449
+▁things	-7.44158
+▁1.	-7.44277
+▁S	-7.45003
+▁then	-7.45549
+▁got	-7.45696
+ur	-7.46269
+▁she	-7.46344
+▁into	-7.47184
+▁Thank	-7.48229
+▁day	-7.48319
+▁very	-7.49401
+ion	-7.49725
+▁provide	-7.49999
+▁try	-7.50489
+▁anything	-7.517
+▁glad	-7.52529
+k	-7.52825
+▁He	-7.54209
+▁him	-7.54619
+u	-7.54986
+ar	-7.56208
+b	-7.57027
+▁life	-7.57379
+▁let	-7.58057
+▁over	-7.58364
+▁why	-7.58373
+▁because	-7.58704
+ation	-7.58798
+▁definitely	-7.59938
+▁back	-7.60564
+▁didn	-7.60604
+ic	-7.6089
+▁C	-7.61403
+▁find	-7.61675
+▁happy	-7.61853
+▁My	-7.62429
+▁never	-7.6253
+▁give	-7.6259
+la	-7.62832
+li	-7.62855
+ra	-7.63821
+▁being	-7.64537
+▁fun	-7.64548
+ne	-7.64712
+ch	-7.65079
+▁best	-7.66179
+▁better	-7.66351
+A	-7.66579
+▁They	-7.66849
+h	-7.67007
+▁look	-7.67069
+▁hard	-7.67201
+on	-7.67574
+▁B	-7.68683
+▁today	-7.69257
+▁use	-7.69921
+ri	-7.70267
+▁someone	-7.70634
+▁first	-7.7113
+▁than	-7.71633
+▁course	-7.72763
+ers	-7.73194
+▁information	-7.73748
+▁many	-7.73971
+ce	-7.74271
+▁come	-7.74741
+▁start	-7.74814
+▁If	-7.74827
+▁home	-7.76898
+▁There	-7.76931
+)	-7.77131
+f	-7.77185
+se	-7.77419
+T	-7.7794
+el	-7.77979
+▁where	-7.78052
+▁still	-7.7895
+▁Here	-7.79841
+C	-7.79973
+▁P	-7.80008
+ter	-7.80622
+▁different	-7.81016
+▁kind	-7.81038
+▁down	-7.81215
+▁few	-7.81408
+▁little	-7.81527
+▁said	-7.81527
+▁hear	-7.81634
+;	-7.81696
+▁job	-7.81828
+ro	-7.81884
+▁does	-7.81907
+D	-7.81974
+▁E	-7.83003
+able	-7.83097
+▁man	-7.83207
+▁only	-7.83942
+▁pretty	-7.83996
+▁F	-7.84347
+▁guess	-7.8488
+me	-7.85391
+▁Let	-7.85705
+▁even	-7.86084
+us	-7.86445
+▁thing	-7.88791
+▁talk	-7.88854
+te	-7.89086
+▁understand	-7.8916
+▁Of	-7.89882
+▁idea	-7.89977
+est	-7.90402
+na	-7.90755
+▁problem	-7.91348
+▁car	-7.91425
+▁made	-7.91472
+M	-7.91656
+▁everything	-7.92032
+▁wanted	-7.92323
+▁okay	-7.92508
+▁long	-7.9279
+▁around	-7.92892
+▁most	-7.92951
+▁Okay	-7.93036
+▁friends	-7.93496
+lo	-7.94228
+▁bad	-7.94351
+▁off	-7.94385
+▁(	-7.94389
+▁In	-7.9495
+▁please	-7.94975
+▁feeling	-7.95416
+▁also	-7.95466
+▁might	-7.96342
+▁getting	-7.96785
+▁Not	-7.96895
+▁trying	-7.97024
+▁thinking	-7.97747
+ir	-7.97949
+P	-7.98492
+▁keep	-7.99194
+il	-7.99448
+▁believe	-7.99481
+▁“	-8.0052
+▁appreciate	-8.00821
+▁looking	-8.01421
+▁To	-8.01462
+▁amazing	-8.01603
+▁Are	-8.02199
+▁world	-8.02301
+▁lately	-8.02322
+▁used	-8.02978
+de	-8.03305
+▁Wow	-8.03462
+▁before	-8.03505
+▁through	-8.04648
+/	-8.04684
+▁school	-8.05096
+▁nice	-8.05683
+▁person	-8.06099
+ist	-8.06556
+ta	-8.06659
+▁those	-8.0694
+▁these	-8.07063
+▁money	-8.07088
+▁while	-8.07153
+▁again	-8.07333
+ry	-8.07535
+ent	-8.07663
+▁such	-8.07845
+E	-8.07936
+O	-8.08161
+ment	-8.08457
+as	-8.08617
+▁us	-8.08617
+to	-8.09341
+po	-8.09365
+ke	-8.09385
+B	-8.09772
+ma	-8.09861
+▁after	-8.10176
+is	-8.10264
+▁big	-8.10469
+▁Hi	-8.10559
+▁Because	-8.11147
+▁play	-8.11261
+▁1	-8.1175
+▁de	-8.12089
+no	-8.12322
+vi	-8.12436
+ge	-8.13102
+▁else	-8.13133
+▁place	-8.13465
+sh	-8.13534
+et	-8.13619
+id	-8.1384
+▁doesn	-8.13947
+▁actually	-8.14612
+▁same	-8.14901
+▁examples	-8.15059
+age	-8.15111
+▁call	-8.15499
+▁f	-8.15604
+ate	-8.15797
+pe	-8.15888
+v	-8.16224
+']	-8.17591
+▁put	-8.17717
+ive	-8.17778
+▁care	-8.18188
+▁add	-8.18434
+▁un	-8.18815
+ity	-8.19295
+▁myself	-8.19548
+▁ever	-8.1978
+▁makes	-8.1998
+ck	-8.20341
+▁specific	-8.20671
+▁two	-8.21248
+▁Ma	-8.21294
+▁last	-8.21389
+▁name	-8.22738
+ol	-8.22779
+▁bit	-8.23142
+▁must	-8.23261
+▁Me	-8.23311
+▁thank	-8.23475
+”	-8.23637
+▁which	-8.24101
+▁working	-8.24533
+ti	-8.24585
+mo	-8.24707
+▁together	-8.24752
+▁years	-8.24841
+▁favorite	-8.25347
+▁She	-8.25688
+hi	-8.26117
+ting	-8.26315
+at	-8.26438
+▁done	-8.2671
+▁own	-8.27871
+▁old	-8.28136
+▁2	-8.2843
+▁show	-8.29045
+z	-8.29775
+▁family	-8.30171
+▁O	-8.30689
+▁Write	-8.30729
+▁hope	-8.30989
+▁true	-8.31082
+L	-8.31385
+ca	-8.31449
+ie	-8.32252
+less	-8.32269
+▁story	-8.32344
+▁T	-8.32799
+▁friend	-8.33272
+▁important	-8.33931
+▁b	-8.34089
+▁Certainly	-8.3422
+▁maybe	-8.34625
+▁wrong	-8.34729
+▁Have	-8.35115
+▁yourself	-8.3546
+▁3	-8.35884
+▁question	-8.35971
+▁tough	-8.3601
+▁part	-8.36029
+▁won	-8.36147
+▁having	-8.36345
+▁making	-8.36348
+ine	-8.36442
+▁food	-8.36568
+▁since	-8.37537
+nt	-8.38346
+▁next	-8.38667
+x	-8.38742
+▁interesting	-8.39055
+pa	-8.39101
+▁without	-8.39138
+ci	-8.3915
+F	-8.39267
+▁talking	-8.39696
+▁live	-8.39945
+▁tips	-8.40044
+▁using	-8.40062
+co	-8.40375
+am	-8.41256
+▁enjoy	-8.41518
+▁nothing	-8.4167
+ba	-8.41716
+▁happened	-8.41893
+um	-8.42216
+H	-8.42231
+▁year	-8.42239
+▁music	-8.42283
+G	-8.42532
+▁end	-8.42662
+▁away	-8.42784
+bo	-8.42839
+ni	-8.42866
+8	-8.42986
+▁everyone	-8.43251
+▁enough	-8.4348
+▁Re	-8.44117
+▁Maybe	-8.44362
+▁mind	-8.44374
+▁high	-8.44517
+▁fine	-8.44565
+▁yet	-8.44834
+▁though	-8.44919
+▁every	-8.45715
+ful	-8.46189
+▁house	-8.4626
+▁D	-8.46609
+▁stay	-8.46622
+ty	-8.46624
+▁each	-8.47389
+▁When	-8.47645
+▁wish	-8.47727
+▁As	-8.47761
+▁M	-8.4779
+▁point	-8.47946
+R	-8.48093
+▁ask	-8.4887
+▁able	-8.49178
+▁Did	-8.49229
+▁Absolutely	-8.49713
+ies	-8.50218
+up	-8.50515
+▁sometimes	-8.50618
+da	-8.51091
+▁game	-8.51146
+▁night	-8.51206
+▁&	-8.5126
+▁looks	-8.51315
+ul	-8.51393
+▁hurt	-8.51422
+▁support	-8.51508
+ho	-8.52766
+▁cool	-8.53356
+man	-8.53385
+▁both	-8.53617
+▁["	-8.5363
+▁change	-8.5385
+ha	-8.53926
+▁found	-8.54221
+▁experience	-8.54658
+ia	-8.54857
+op	-8.55092
+ow	-8.55148
+▁anyone	-8.55326
+U	-8.56116
+land	-8.56561
+cu	-8.56586
+▁far	-8.56869
+W	-8.56919
+ru	-8.57056
+be	-8.57095
+▁Don	-8.5744
+▁text	-8.57459
+un	-8.57939
+▁seem	-8.58068
+▁sound	-8.58105
+▁H	-8.58292
+▁De	-8.58294
+▁real	-8.5868
+ad	-8.58772
+fe	-8.59132
+N	-8.5964
+▁taking	-8.59771
+▁went	-8.5979
+▁Trump	-8.59823
+ut	-8.60113
+2	-8.60421
+im	-8.60695
+bu	-8.61096
+▁stop	-8.61558
+▁Ha	-8.61689
+▁isn	-8.61794
+do	-8.61893
+▁learn	-8.62303
+▁suggest	-8.6242
+▁K	-8.62461
+▁wait	-8.62519
+ten	-8.62552
+▁head	-8.62609
+▁heard	-8.62661
+Y	-8.62698
+per	-8.62834
+▁between	-8.63194
+▁For	-8.63275
+▁team	-8.63446
+▁Where	-8.63874
+▁sense	-8.64251
+▁read	-8.64318
+bi	-8.64443
+▁during	-8.64659
+king	-8.64902
+ness	-8.64925
+▁busy	-8.65115
+▁Mom	-8.65912
+▁Sarah	-8.66142
+▁thanks	-8.6619
+▁means	-8.66776
+he	-8.66863
+▁stuff	-8.66945
+>	-8.6765
+▁worry	-8.67954
+▁W	-8.68014
+▁...	-8.6811
+▁eat	-8.68169
+by	-8.68233
+▁remember	-8.68565
+▁program	-8.68873
+▁G	-8.68941
+▁week	-8.69409
+▁small	-8.69699
+▁guy	-8.69789
+▁Who	-8.70001
+▁already	-8.70318
+▁check	-8.70425
+▁<	-8.70778
+▁An	-8.71129
+▁leave	-8.71154
+▁probably	-8.71363
+▁coming	-8.71418
+ot	-8.71422
+▁happen	-8.7161
+▁deal	-8.71791
+▁Be	-8.71939
+▁seen	-8.71942
+son	-8.71961
+ber	-8.72073
+▁impact	-8.72315
+▁another	-8.72739
+va	-8.72983
+▁break	-8.73165
+▁haven	-8.73437
+▁asking	-8.73505
+▁Good	-8.73601
+▁details	-8.73636
+▁easy	-8.73703
+ver	-8.73838
+om	-8.74563
+▁number	-8.74676
+and	-8.74689
+nd	-8.74787
+▁may	-8.7508
+▁soon	-8.75341
+▁beautiful	-8.75439
+▁set	-8.75828
+▁started	-8.75946
+▁recipe	-8.76098
+▁store	-8.76335
+▁test	-8.76402
+ance	-8.76494
+▁pro	-8.76586
+▁Your	-8.76962
+▁Great	-8.77318
+▁saw	-8.77581
+▁sta	-8.77688
+]	-8.77716
+ish	-8.77832
+▁called	-8.77998
+▁meet	-8.7811
+▁seems	-8.7847
+▁came	-8.78631
+mon	-8.78707
+▁dis	-8.78735
+▁told	-8.79007
+▁con	-8.79051
+5	-8.79256
+9	-8.79272
+▁several	-8.79346
+dy	-8.7954
+ee	-8.79655
+fi	-8.79685
+▁project	-8.80146
+ph	-8.80198
+▁Ja	-8.80311
+▁water	-8.80342
+1	-8.80561
+▁explain	-8.80721
+3	-8.80736
+▁free	-8.80748
+▁L	-8.80947
+▁plan	-8.81293
+▁dog	-8.81503
+tic	-8.81619
+▁Po	-8.8172
+▁social	-8.81767
+ard	-8.81991
+▁class	-8.82212
+line	-8.82262
+▁agree	-8.82291
+lu	-8.82407
+▁matter	-8.82483
+▁Now	-8.82795
+▁kids	-8.82898
+▁book	-8.83087
+▁N	-8.83294
+▁close	-8.83301
+▁Co	-8.83338
+▁answer	-8.84006
+▁Could	-8.84006
+▁4	-8.84063
+▁buy	-8.84126
+▁short	-8.84126
+▁hanging	-8.84383
+▁country	-8.84497
+▁welcome	-8.84724
+▁business	-8.84884
+em	-8.84984
+▁hate	-8.84993
+▁post	-8.85319
+▁wife	-8.85462
+▁ways	-8.85487
+go	-8.85666
+▁Na	-8.86111
+▁least	-8.8642
+URL	-8.86423
+▁city	-8.86569
+ice	-8.86742
+▁worth	-8.86787
+▁La	-8.86932
+▁black	-8.87284
+▁pi	-8.8735
+▁bar	-8.87508
+▁watch	-8.87543
+▁pay	-8.87727
+che	-8.87737
+▁company	-8.87897
+▁under	-8.88681
+tor	-8.89166
+▁left	-8.89275
+▁woman	-8.89399
+▁perfect	-8.89638
+▁gra	-8.89939
+4	-8.90013
+▁recommend	-8.90026
+ian	-8.90032
+▁run	-8.90091
+▁finally	-8.90183
+▁mom	-8.90217
+ous	-8.90282
+▁says	-8.90479
+rn	-8.90482
+▁interested	-8.90519
+▁feels	-8.90679
+ating	-8.90866
+▁cat	-8.90932
+act	-8.91067
+▁ideas	-8.91107
+tra	-8.91551
+les	-8.91616
+▁excited	-8.91688
+▁list	-8.91761
+ted	-8.91884
+▁open	-8.92094
+▁imagine	-8.92204
+▁ready	-8.92719
+▁pre	-8.92954
+▁Ba	-8.93004
+▁co	-8.93029
+ally	-8.93115
+▁Ro	-8.93271
+▁All	-8.93344
+▁figure	-8.93901
+▁playing	-8.94411
+▁face	-8.94546
+▁address	-8.94697
+str	-8.95028
+▁girl	-8.95479
+▁One	-8.95678
+▁health	-8.95707
+▁offer	-8.95721
+cr	-8.95767
+io	-8.96039
+▁R	-8.96063
+rs	-8.96249
+K	-8.9628
+▁$	-8.96445
+▁5	-8.9669
+▁hit	-8.96741
+vo	-8.96877
+▁color	-8.96989
+▁party	-8.97184
+▁wasn	-8.97231
+ag	-8.97269
+▁U	-8.97437
+▁Mar	-8.97493
+fa	-8.97592
+▁Please	-8.97795
+▁Li	-8.98038
+*	-8.98057
+▁days	-8.98222
+▁personal	-8.98432
+▁video	-8.98692
+di	-8.98706
+ton	-8.98847
+du	-8.99338
+▁options	-8.99577
+V	-8.99815
+▁access	-8.99845
+ac	-8.99865
+▁move	-8.99877
+ated	-8.99953
+▁pick	-8.99961
+▁ex	-8.99986
+▁women	-9.0015
+ight	-9.00412
+side	-9.00447
+▁hand	-9.01062
+▁phone	-9.01249
+▁saying	-9.01427
+▁lost	-9.01431
+ner	-9.01444
+▁case	-9.01444
+ga	-9.01614
+wa	-9.01614
+▁living	-9.01703
+way	-9.01711
+ial	-9.01909
+▁fair	-9.02005
+ling	-9.02119
+▁paper	-9.02136
+▁create	-9.02383
+▁Th	-9.02498
+▁design	-9.02688
+▁parents	-9.03135
+▁human	-9.0329
+ick	-9.0367
+▁20	-9.03993
+▁history	-9.04029
+▁once	-9.04182
+▁art	-9.04653
+we	-9.04754
+pp	-9.04774
+▁Ra	-9.04936
+▁kill	-9.04943
+▁difference	-9.05033
+ever	-9.05037
+fo	-9.05172
+the	-9.05254
+▁against	-9.0541
+▁10	-9.0545
+▁become	-9.05772
+▁movie	-9.06082
+hy	-9.06092
+rt	-9.06104
+mer	-9.06369
+▁New	-9.06403
+▁Se	-9.06449
+▁morning	-9.06664
+▁quite	-9.06707
+▁shit	-9.06713
+▁Or	-9.06874
+ea	-9.06946
+▁enjoying	-9.06963
+▁system	-9.07114
+▁possible	-9.07264
+▁step	-9.074
+▁turn	-9.07407
+▁trouble	-9.07415
+▁tried	-9.07416
+tro	-9.07513
+▁Mo	-9.07551
+ger	-9.07559
+▁Sa	-9.07621
+▁follow	-9.0778
+▁common	-9.08032
+ary	-9.08101
+▁room	-9.08258
+▁sex	-9.08302
+ize	-9.08307
+▁anymore	-9.08324
+The	-9.08389
+ud	-9.0891
+",	-9.08985
+▁drink	-9.09002
+▁white	-9.09136
+ig	-9.09199
+que	-9.09248
+▁Really	-9.0942
+▁Ka	-9.09501
+ism	-9.09568
+tu	-9.0969
+ction	-9.0969
+▁material	-9.09701
+▁God	-9.09821
+▁group	-9.09856
+▁safe	-9.10015
+▁piece	-9.1002
+der	-9.10376
+▁including	-9.10631
+▁St	-9.10651
+time	-9.10965
+ng	-9.11471
+ship	-9.11611
+day	-9.11795
+▁couldn	-9.12026
+▁focus	-9.12199
+7	-9.12428
+ap	-9.12619
+▁wondering	-9.12725
+▁reason	-9.12735
+▁light	-9.12869
+▁words	-9.12912
+▁doctor	-9.12993
+port	-9.13189
+▁proud	-9.1333
+▁exactly	-9.13338
+▁Ca	-9.13385
+ence	-9.13825
+mar	-9.14114
+▁heart	-9.14255
+▁promise	-9.14276
+▁asked	-9.14382
+end	-9.1442
+▁walk	-9.14505
+j	-9.14513
+▁Ho	-9.14769
+▁chance	-9.14934
+▁wouldn	-9.1497
+▁travel	-9.15087
+▁Al	-9.15109
+ny	-9.15287
+▁process	-9.15413
+▁online	-9.1543
+▁*	-9.15483
+rr	-9.15759
+ring	-9.15765
+▁V	-9.15814
+low	-9.15943
+▁full	-9.15979
+wo	-9.16193
+ok	-9.16243
+▁win	-9.1645
+▁stand	-9.1679
+▁Bo	-9.1695
+▁past	-9.17076
+▁whole	-9.17171
+▁research	-9.17251
+all	-9.1729
+▁helpful	-9.17314
+ah	-9.17316
+▁young	-9.17613
+▁media	-9.17662
+▁trip	-9.17733
+▁until	-9.17899
+con	-9.1794
+out	-9.17947
+▁cook	-9.18024
+▁per	-9.18168
+gg	-9.18277
+▁wonder	-9.18483
+▁son	-9.18716
+▁often	-9.18716
+▁future	-9.18733
+▁Then	-9.18769
+▁children	-9.18902
+▁Pro	-9.18906
+ster	-9.18974
+▁clean	-9.19071
+▁line	-9.19135
+oo	-9.19206
+▁Te	-9.19266
+▁sub	-9.19291
+▁cl	-9.19369
+X	-9.1941
+▁boss	-9.19438
+amp	-9.19593
+▁three	-9.20093
+▁miss	-9.20117
+▁tonight	-9.20153
+▁speak	-9.20216
+▁times	-9.20332
+▁special	-9.2038
+▁baby	-9.20612
+▁war	-9.20618
+fl	-9.20711
+▁side	-9.2072
+▁app	-9.20892
+▁order	-9.21129
+od	-9.21313
+▁type	-9.2132
+os	-9.21475
+▁wi	-9.21486
+rate	-9.2154
+tion	-9.2154
+▁difficult	-9.21644
+ical	-9.21735
+ey	-9.21739
+▁visit	-9.22125
+▁goes	-9.22174
+▁funny	-9.22217
+▁took	-9.2224
+▁im	-9.22399
+▁fan	-9.22494
+▁American	-9.22501
+▁Le	-9.22936
+▁Da	-9.23252
+▁sun	-9.23288
+one	-9.23459
+oh	-9.23463
+ect	-9.2368
+6	-9.23927
+▁mother	-9.23931
+▁news	-9.2403
+▁sweet	-9.24084
+rap	-9.24088
+my	-9.2411
+▁ago	-9.24146
+▁careful	-9.24236
+▁comp	-9.24277
+▁language	-9.24284
+▁alone	-9.24483
+av	-9.24567
+▁yeah	-9.2465
+▁ha	-9.24676
+ka	-9.24743
+▁popular	-9.24821
+▁yes	-9.24975
+ep	-9.25157
+ach	-9.25165
+our	-9.25169
+▁either	-9.25181
+iv	-9.25255
+▁Pa	-9.25272
+▁word	-9.25356
+com	-9.25796
+▁college	-9.25903
+▁starting	-9.26219
+▁John	-9.26251
+lar	-9.26261
+▁body	-9.26281
+▁positive	-9.26449
+▁games	-9.26531
+pl	-9.26611
+ries	-9.26674
+▁guys	-9.26743
+▁sick	-9.26795
+▁strong	-9.26815
+ise	-9.26945
+▁worried	-9.26997
+▁thou	-9.27365
+par	-9.27388
+▁Mi	-9.27542
+▁decision	-9.27549
+ble	-9.27595
+▁given	-9.2762
+▁Would	-9.27733
+▁share	-9.28008
+▁fire	-9.28053
+▁improve	-9.28103
+▁website	-9.28103
+▁men	-9.28291
+▁tired	-9.28477
+▁Alright	-9.28478
+▁state	-9.2875
+▁situation	-9.29034
+▁hair	-9.29083
+▁Which	-9.29222
+▁user	-9.29273
+▁[	-9.29451
+▁dead	-9.2948
+▁comes	-9.29585
+▁power	-9.29587
+▁thy	-9.29963
+▁relax	-9.30151
+▁fall	-9.30244
+▁self	-9.30392
+▁dream	-9.30543
+▁Some	-9.30693
+▁park	-9.3083
+▁write	-9.30887
+▁Je	-9.31028
+▁outside	-9.31167
+▁6	-9.31533
+▁report	-9.31573
+▁Su	-9.31865
+▁marketing	-9.32198
+▁ho	-9.32203
+▁fact	-9.32221
+ear	-9.32403
+ab	-9.32464
+▁Go	-9.32549
+▁Di	-9.32594
+ak	-9.32751
+▁law	-9.32952
+▁advice	-9.33042
+00	-9.33068
+▁listen	-9.33142
+▁large	-9.33228
+▁child	-9.33235
+bs	-9.33368
+▁Con	-9.33977
+▁relationship	-9.34021
+▁J	-9.3423
+▁Dad	-9.34416
+ture	-9.34444
+▁Jo	-9.34563
+wi	-9.34587
+▁air	-9.34591
+▁Like	-9.34658
+ps	-9.3473
+▁With	-9.3484
+▁US	-9.35091
+▁rest	-9.35183
+▁deserve	-9.35208
+chi	-9.35277
+ex	-9.3535
+▁tomorrow	-9.35459
+qui	-9.35583
+▁types	-9.35825
+ya	-9.35853
+▁Hmm	-9.36022
+▁father	-9.36024
+▁These	-9.3624
+unt	-9.3633
+▁example	-9.36683
+▁kid	-9.36742
+ron	-9.36752
+▁cup	-9.37294
+ite	-9.37384
+ub	-9.37391
+▁books	-9.37452
+▁sign	-9.37532
+▁forward	-9.37546
+▁unique	-9.37625
+▁Pre	-9.37826
+▁second	-9.38033
+led	-9.38169
+ative	-9.38204
+(	-9.38236
+▁technology	-9.38238
+▁19	-9.38344
+ping	-9.38443
+▁mine	-9.3854
+▁sp	-9.3856
+▁control	-9.3865
+▁card	-9.38774
+▁gone	-9.38822
+bl	-9.38909
+_	-9.39062
+ip	-9.39133
+ob	-9.39309
+ition	-9.39403
+▁guide	-9.39689
+▁public	-9.39772
+▁cause	-9.39783
+▁watching	-9.39933
+dditional	-9.40113
+▁save	-9.40181
+▁late	-9.40299
+▁OK	-9.40357
+▁drive	-9.40542
+▁Z	-9.40583
+▁role	-9.40605
+▁spend	-9.40775
+▁virus	-9.40942
+▁moment	-9.41022
+▁space	-9.41082
+ious	-9.4114
+▁felt	-9.41157
+ff	-9.41249
+▁data	-9.41376
+▁alright	-9.41794
+▁learning	-9.41803
+den	-9.41827
+cent	-9.41962
+ped	-9.42112
+ure	-9.42186
+▁culture	-9.42216
+ja	-9.42348
+▁chicken	-9.42427
+▁dinner	-9.42428
+▁key	-9.42534
+ang	-9.42583
+0	-9.42647
+▁brother	-9.42921
+▁significant	-9.43071
+▁reading	-9.43101
+▁fight	-9.43163
+bb	-9.43224
+J	-9.43272
+▁fast	-9.43362
+▁awesome	-9.43503
+▁town	-9.43509
+▁Ar	-9.43603
+ome	-9.43631
+▁Ta	-9.43836
+▁sleep	-9.43981
+▁leg	-9.44065
+▁available	-9.44153
+▁potential	-9.44153
+▁local	-9.44157
+▁date	-9.44168
+▁forget	-9.44177
+▁loved	-9.44192
+▁instead	-9.44371
+▁knew	-9.44371
+▁season	-9.4438
+▁Ga	-9.44477
+▁upset	-9.44616
+▁Lo	-9.44785
+ther	-9.44807
+cy	-9.44954
+nu	-9.45058
+lin	-9.45205
+▁cut	-9.45233
+▁cold	-9.45326
+▁weekend	-9.45491
+sha	-9.4554
+▁product	-9.45545
+▁pass	-9.45774
+ward	-9.45871
+ki	-9.46086
+▁running	-9.46146
+cur	-9.46165
+▁arm	-9.46186
+▁crazy	-9.46354
+▁teacher	-9.46364
+board	-9.46397
+▁Du	-9.46442
+▁boy	-9.46649
+▁hot	-9.46673
+▁months	-9.46997
+▁pain	-9.47043
+▁email	-9.4726
+▁gun	-9.47345
+▁area	-9.47448
+ue	-9.47507
+▁successful	-9.47575
+▁afraid	-9.47698
+▁7	-9.47861
+ning	-9.47923
+▁Ju	-9.4804
+▁Create	-9.48146
+▁price	-9.48165
+▁trust	-9.48409
+gi	-9.48508
+▁detailed	-9.48526
+press	-9.48589
+▁wear	-9.48723
+▁top	-9.48759
+▁article	-9.48832
+▁police	-9.48834
+▁caught	-9.48835
+▁die	-9.48983
+▁behind	-9.49061
+ase	-9.49377
+▁fuck	-9.49392
+▁based	-9.49456
+▁death	-9.49541
+▁coffee	-9.49748
+▁energy	-9.49748
+▁industry	-9.49748
+▁di	-9.5013
+▁usual	-9.50337
+use	-9.50343
+▁Exactly	-9.50441
+ose	-9.50471
+lt	-9.50668
+▁suggestions	-9.50677
+▁computer	-9.50905
+▁smart	-9.50924
+▁Tell	-9.5094
+▁shall	-9.50955
+▁service	-9.50972
+▁healthy	-9.51461
+▁usually	-9.5151
+word	-9.51734
+▁Dr	-9.51775
+yn	-9.51863
+ible	-9.51932
+▁animals	-9.51943
+ding	-9.51981
+▁weeks	-9.52006
+▁looked	-9.52128
+▁song	-9.52193
+▁fr	-9.52232
+step	-9.52321
+lay	-9.52525
+▁inside	-9.52592
+▁market	-9.52635
+▁Come	-9.52723
+▁view	-9.52757
+▁needed	-9.52778
+old	-9.52949
+lf	-9.52997
+▁main	-9.5301
+▁hours	-9.53055
+▁government	-9.53262
+▁mi	-9.53473
+ami	-9.53498
+▁poem	-9.53509
+rac	-9.53514
+ier	-9.53612
+▁sort	-9.53691
+▁develop	-9.53698
+▁worked	-9.53776
+▁dance	-9.53839
+▁building	-9.54025
+▁red	-9.54069
+ile	-9.54082
+pri	-9.54092
+▁street	-9.54219
+▁general	-9.54221
+ail	-9.54321
+▁lives	-9.54349
+▁sit	-9.54364
+lock	-9.54649
+ler	-9.5465
+▁hands	-9.54819
+ley	-9.54889
+men	-9.54893
+▁exciting	-9.54944
+▁giving	-9.55113
+▁film	-9.55188
+▁places	-9.55209
+▁eyes	-9.55301
+Q	-9.55389
+▁mistake	-9.55431
+▁happens	-9.55446
+und	-9.55521
+▁tree	-9.55542
+▁scared	-9.55778
+▁effective	-9.55791
+▁almost	-9.55918
+▁aren	-9.56084
+▁writing	-9.56163
+▁sister	-9.56172
+▁Fa	-9.56385
+uck	-9.56395
+▁supposed	-9.56514
+ability	-9.56686
+▁near	-9.5671
+30	-9.56766
+▁clear	-9.56779
+q	-9.5687
+▁toilet	-9.56923
+▁catch	-9.5709
+▁Sounds	-9.57154
+ric	-9.57406
+▁pet	-9.57463
+▁sad	-9.57539
+if	-9.57553
+▁serious	-9.57761
+▁fit	-9.57918
+▁bring	-9.58005
+bra	-9.58082
+tain	-9.58298
+form	-9.58459
+▁steal	-9.58511
+ju	-9.5863
+▁benefits	-9.58836
+▁cons	-9.58844
+▁takes	-9.58939
+cover	-9.5897
+▁walking	-9.59003
+▁At	-9.59013
+▁Mr	-9.59193
+ld	-9.59218
+-19	-9.59262
+▁America	-9.59284
+That	-9.59379
+▁Un	-9.59394
+▁Ad	-9.59454
+ized	-9.5948
+cha	-9.59494
+▁Fine	-9.59559
+▁lucky	-9.59568
+▁Sorry	-9.5966
+50	-9.59685
+wn	-9.59719
+▁cost	-9.59749
+▁husband	-9.59914
+▁opinion	-9.59914
+▁study	-9.60034
+▁decided	-9.60215
+oc	-9.60577
+▁explore	-9.60682
+pu	-9.60738
+pen	-9.60842
+ai	-9.61164
+▁seeing	-9.61313
+▁whatever	-9.6136
+▁fault	-9.61454
+▁choice	-9.61456
+▁mental	-9.61607
+▁natural	-9.61713
+▁essay	-9.61973
+fu	-9.61999
+ris	-9.6216
+▁bank	-9.62228
+▁couple	-9.62233
+▁eye	-9.62253
+▁touch	-9.62374
+▁extra	-9.62549
+▁Hu	-9.62581
+gan	-9.6263
+▁Sha	-9.62667
+ors	-9.6276
+▁especially	-9.62761
+▁Car	-9.62971
+ze	-9.63347
+art	-9.63523
+▁restaurant	-9.63546
+▁track	-9.6356
+val	-9.63563
+wood	-9.63567
+ual	-9.63581
+ack	-9.6359
+▁helping	-9.6366
+▁Vi	-9.63882
+▁weather	-9.6392
+▁handle	-9.63977
+▁practice	-9.64033
+▁character	-9.64058
+box	-9.64256
+▁Ki	-9.64413
+▁Does	-9.646
+▁delicious	-9.64608
+▁hoping	-9.6466
+▁lie	-9.64734
+▁join	-9.64887
+▁Man	-9.64919
+rk	-9.6502
+ew	-9.65127
+▁ball	-9.65185
+that	-9.65202
+lia	-9.65223
+nn	-9.65288
+▁TV	-9.65472
+ator	-9.65637
+▁stupid	-9.65682
+▁account	-9.65931
+▁lol	-9.66038
+▁Gr	-9.66116
+su	-9.66209
+▁blood	-9.66224
+▁various	-9.66228
+▁telling	-9.6628
+▁letter	-9.66322
+ple	-9.66439
+15	-9.6644
+▁decide	-9.66451
+▁month	-9.6655
+▁works	-9.66604
+ump	-9.66754
+▁terrible	-9.66768
+ink	-9.66864
+▁honest	-9.66905
+▁level	-9.67041
+▁dark	-9.67044
+▁angry	-9.67048
+▁later	-9.67243
+?"	-9.67309
+elect	-9.67313
+▁English	-9.67315
+▁physical	-9.67315
+▁laugh	-9.67321
+▁rock	-9.67339
+▁Che	-9.67413
+▁office	-9.67415
+ay	-9.67496
+▁exercise	-9.67589
+▁fresh	-9.6759
+ug	-9.67591
+cho	-9.6777
+ina	-9.67826
+▁People	-9.67865
+▁Ex	-9.67995
+au	-9.68001
+▁trans	-9.68119
+▁grow	-9.68145
+gen	-9.68414
+▁daughter	-9.68418
+▁major	-9.68418
+Oh	-9.68532
+My	-9.68764
+▁breath	-9.68975
+▁accept	-9.68976
+200	-9.6904
+▁current	-9.69118
+▁door	-9.69171
+hu	-9.69213
+▁married	-9.69254
+▁picture	-9.69255
+▁math	-9.69446
+▁girlfriend	-9.69537
+▁known	-9.69545
+▁include	-9.69699
+▁plant	-9.70013
+▁community	-9.70105
+20	-9.70278
+ign	-9.70293
+▁throw	-9.70387
+▁cake	-9.70396
+imp	-9.70416
+▁deep	-9.70452
+pack	-9.70582
+▁comprehensive	-9.70662
+▁four	-9.70733
+▁particular	-9.70946
+mate	-9.70998
+ray	-9.71051
+▁Ri	-9.71181
+friendly	-9.71347
+▁hold	-9.71379
+▁consider	-9.71394
+▁cap	-9.71553
+▁traditional	-9.71666
+AN	-9.71803
+▁value	-9.71806
+▁El	-9.71841
+range	-9.72041
+▁search	-9.72091
+▁early	-9.7212
+▁recently	-9.722
+▁single	-9.7241
+▁luck	-9.72504
+▁action	-9.72782
+ide	-9.72903
+▁lose	-9.72926
+▁avoid	-9.72959
+▁rather	-9.72975
+%	-9.73108
+▁fix	-9.73289
+wr	-9.73312
+boo	-9.73481
+▁bus	-9.73521
+▁cooking	-9.73636
+well	-9.7367
+▁Ne	-9.73838
+▁minutes	-9.73847
+▁band	-9.73926
+lie	-9.74065
+▁due	-9.74086
+▁named	-9.74117
+eg	-9.74268
+▁meeting	-9.74278
+▁clothes	-9.74715
+nce	-9.74773
+▁completely	-9.748
+▁movies	-9.74982
+▁shouldn	-9.75068
+▁super	-9.75138
+br	-9.75205
+▁pea	-9.75253
+▁instructions	-9.75268
+▁crime	-9.75322
+▁shot	-9.75335
+▁development	-9.75359
+▁count	-9.75361
+▁War	-9.75895
+▁describe	-9.76203
+▁contact	-9.76203
+▁ran	-9.76326
+▁send	-9.76352
+▁spending	-9.76467
+▁attack	-9.76506
+nk	-9.76547
+▁steps	-9.76621
+▁Ke	-9.76706
+▁brand	-9.7681
+▁nature	-9.76813
+▁taste	-9.7686
+12	-9.76915
+log	-9.76942
+▁anti	-9.76948
+▁Look	-9.7698
+▁incredible	-9.77107
+RE	-9.77108
+cal	-9.77227
+▁customer	-9.77258
+▁His	-9.77342
+▁dogs	-9.77364
+▁version	-9.77411
+▁content	-9.77411
+▁along	-9.77421
+▁talked	-9.77914
+pic	-9.77945
+▁Chinese	-9.78019
+▁gonna	-9.7802
+▁smell	-9.78027
+▁certainly	-9.78121
+▁afford	-9.78325
+▁affect	-9.78325
+gue	-9.78389
+▁Ve	-9.78404
+aw	-9.78448
+▁sell	-9.7849
+ix	-9.78519
+▁mentioned	-9.78531
+van	-9.78587
+▁fascinating	-9.78631
+▁simple	-9.78634
+ware	-9.78678
+een	-9.78769
+▁dress	-9.78911
+▁countries	-9.78948
+▁issue	-9.79133
+▁driving	-9.79247
+▁lead	-9.79404
+▁Any	-9.79414
+zz	-9.79495
+▁others	-9.79498
+▁students	-9.79531
+▁COVID	-9.79557
+▁effort	-9.79557
+▁remind	-9.79566
+▁field	-9.79613
+▁On	-9.7967
+ft	-9.79845
+▁Our	-9.79859
+cing	-9.79938
+▁fill	-9.79949
+▁meal	-9.7999
+ye	-9.79997
+▁rep	-9.80125
+▁birthday	-9.80183
+▁upon	-9.80203
+▁lord	-9.80219
+▁eco	-9.8051
+og	-9.80513
+▁risk	-9.80593
+gar	-9.8079
+▁source	-9.80807
+▁rid	-9.8101
+ank	-9.81086
+ory	-9.81223
+▁suck	-9.81348
+▁skin	-9.8136
+ev	-9.81366
+▁yesterday	-9.81435
+▁First	-9.81435
+▁choose	-9.81436
+▁died	-9.81483
+cast	-9.81523
+**	-9.81538
+▁front	-9.81754
+ncy	-9.81807
+▁table	-9.81866
+▁Sh	-9.81881
+▁mad	-9.81924
+ities	-9.82021
+▁shows	-9.82047
+▁weird	-9.82078
+▁non	-9.82162
+▁normal	-9.82395
+▁eating	-9.82396
+▁planning	-9.8241
+11	-9.8246
+▁fear	-9.8247
+▁played	-9.82504
+cor	-9.82578
+▁entire	-9.82708
+▁discover	-9.82741
+uff	-9.82776
+▁wall	-9.82794
+▁sc	-9.82919
+000	-9.82964
+▁Cr	-9.8297
+ade	-9.82981
+▁meaning	-9.83033
+▁half	-9.83048
+▁Gu	-9.83091
+▁bed	-9.83226
+▁activities	-9.83349
+▁murder	-9.83349
+▁questions	-9.83366
+▁Sta	-9.83382
+▁position	-9.83672
+▁model	-9.83707
+▁understanding	-9.83823
+▁machine	-9.83995
+▁god	-9.83997
+kin	-9.84006
+▁beat	-9.84133
+▁flavor	-9.8432
+▁cannot	-9.84669
+▁forever	-9.84913
+▁shop	-9.84991
+▁hell	-9.84999
+▁expect	-9.85002
+▁serve	-9.85008
+quality	-9.85054
+▁code	-9.85171
+ana	-9.85173
+▁campaign	-9.853
+▁record	-9.85301
+▁message	-9.85332
+▁exist	-9.85383
+mission	-9.85565
++	-9.85629
+▁After	-9.8563
+▁York	-9.85632
+▁finding	-9.85663
+▁rich	-9.85731
+▁artist	-9.85774
+▁race	-9.85946
+▁coronavirus	-9.85959
+▁poor	-9.8602
+▁gave	-9.86032
+▁walks	-9.86054
+▁mess	-9.8612
+▁hospital	-9.8629
+▁author	-9.8629
+▁sir	-9.86549
+▁Will	-9.86569
+▁neighbor	-9.8662
+bal	-9.86673
+ging	-9.86808
+--	-9.86853
+▁Par	-9.86923
+▁accident	-9.86942
+▁allow	-9.86946
+▁application	-9.86956
+▁calm	-9.87105
+▁warm	-9.87139
+▁beach	-9.87169
+ST	-9.87219
+▁return	-9.8729
+▁insult	-9.87293
+▁road	-9.87311
+▁cross	-9.87362
+ER	-9.87555
+pro	-9.87783
+itch	-9.87818
+▁hang	-9.87948
+▁fake	-9.87983
+▁stick	-9.87986
+▁build	-9.88017
+har	-9.88106
+▁meat	-9.8811
+▁noticed	-9.88188
+▁mis	-9.88188
+▁player	-9.88245
+ham	-9.88273
+▁skills	-9.8828
+▁million	-9.88301
+▁achieve	-9.88302
+▁exp	-9.88394
+body	-9.88399
+▁bill	-9.88467
+▁success	-9.88615
+▁worse	-9.8864
+▁draw	-9.88646
+▁charge	-9.88669
+▁finish	-9.88674
+▁Day	-9.88692
+nic	-9.88832
+=	-9.8898
+work	-9.89262
+▁vote	-9.8928
+af	-9.89294
+▁pizza	-9.89321
+▁moving	-9.89321
+row	-9.89322
+▁mask	-9.89323
+▁performance	-9.89336
+▁China	-9.89349
+What	-9.89381
+▁treat	-9.89496
+▁2.	-9.89565
+ON	-9.89569
+▁taken	-9.89584
+▁missed	-9.896
+▁education	-9.89673
+▁raise	-9.89688
+▁green	-9.89688
+▁resources	-9.89818
+lation	-9.89886
+▁garden	-9.90008
+▁anyway	-9.90208
+gl	-9.90306
+▁secret	-9.90352
+▁Ok	-9.90357
+▁May	-9.90427
+No	-9.90617
+▁fish	-9.90961
+▁Ingredients	-9.91045
+▁progress	-9.91049
+▁train	-9.9105
+purpose	-9.9105
+▁push	-9.91061
+▁rain	-9.9121
+what	-9.91391
+▁stuck	-9.914
+▁racist	-9.914
+▁strange	-9.91402
+▁interest	-9.91436
+▁coach	-9.91441
+▁following	-9.91596
+▁honey	-9.91753
+▁events	-9.91817
+AR	-9.91824
+▁career	-9.91893
+▁problems	-9.91897
+▁event	-9.91904
+▁Sometimes	-9.92095
+▁bought	-9.92113
+▁professional	-9.92205
+▁dec	-9.92394
+▁sports	-9.92395
+▁tweet	-9.92449
+▁lunch	-9.92451
+ath	-9.92478
+▁listening	-9.92801
+▁born	-9.92815
+▁willing	-9.92824
+ax	-9.9297
+▁plans	-9.92976
+▁Wa	-9.93107
+▁Nothing	-9.9316
+script	-9.93167
+▁helped	-9.93184
+▁weight	-9.93336
+▁Coronavirus	-9.9351
+app	-9.93696
+▁bag	-9.93935
+read	-9.94176
+▁recent	-9.94227
+▁paid	-9.94245
+▁challenges	-9.94292
+wing	-9.94429
+▁rec	-9.94442
+▁inter	-9.94475
+▁present	-9.94537
+▁Pe	-9.94557
+▁horse	-9.94586
+▁ingredients	-9.94759
+ound	-9.94811
+nder	-9.94957
+ium	-9.95098
+▁studying	-9.95138
+▁Every	-9.95299
+cultural	-9.95309
+▁famous	-9.95309
+▁gu	-9.95326
+▁dishes	-9.95391
+▁gay	-9.95392
+▁protect	-9.95482
+▁continue	-9.95593
+▁United	-9.95673
+▁perform	-9.95675
+▁waiting	-9.95693
+▁pair	-9.95705
+▁hour	-9.9578
+▁worst	-9.96037
+▁grade	-9.96176
+▁ahead	-9.96225
+▁blue	-9.96406
+▁huge	-9.96472
+▁wonderful	-9.96526
+▁Bar	-9.96587
+▁depends	-9.96617
+▁response	-9.96771
+▁scene	-9.96775
+▁totally	-9.96907
+▁dish	-9.96954
+▁features	-9.96984
+bul	-9.97007
+▁Christmas	-9.9714
+▁block	-9.97189
+▁image	-9.97224
+Fi	-9.97329
+igh	-9.97494
+▁chocolate	-9.97511
+▁respect	-9.97511
+lee	-9.97848
+▁method	-9.97883
+▁match	-9.97885
+▁function	-9.98228
+▁pull	-9.98296
+▁Am	-9.98354
+▁pan	-9.98584
+▁across	-9.9863
+▁comfortable	-9.9865
+▁web	-9.98797
+night	-9.98836
+▁businesses	-9.98884
+len	-9.98906
+ugh	-9.98998
+▁plenty	-9.99006
+▁creating	-9.99006
+▁review	-9.99031
+char	-9.99304
+▁cop	-9.99309
+▁See	-9.99314
+ments	-9.99455
+▁By	-9.99517
+▁ski	-9.99557
+▁slow	-9.99609
+istic	-9.99612
+▁roll	-9.99626
+▁Fr	-9.99633
+uc	-9.99826
+▁suit	-9.99879
+▁bear	-9.99892
+▁expensive	-10.0014
+▁illegal	-10.0014
+▁nervous	-10.0014
+▁opportunity	-10.0014
+▁voice	-10.0015
+▁window	-10.0016
+jo	-10.0016
+don	-10.0017
+▁meant	-10.0045
+▁summer	-10.0052
+▁goal	-10.0058
+▁Mu	-10.0069
+▁mix	-10.0069
+▁gas	-10.0074
+▁wearing	-10.0081
+OR	-10.0082
+▁science	-10.0091
+▁Ah	-10.0092
+▁Qu	-10.0097
+▁longer	-10.0098
+print	-10.0099
+ky	-10.0099
+ify	-10.0101
+▁cry	-10.0106
+▁size	-10.0111
+10	-10.0111
+lly	-10.0114
+▁learned	-10.0136
+for	-10.0142
+her	-10.0159
+▁Ru	-10.0166
+▁fruit	-10.0168
+ology	-10.0188
+care	-10.0196
+▁fly	-10.02
+path	-10.0201
+▁bother	-10.0202
+▁limit	-10.0207
+▁milk	-10.0207
+▁software	-10.0207
+▁sometime	-10.023
+like	-10.0239
+▁Everything	-10.0245
+die	-10.0248
+izing	-10.0248
+▁egg	-10.0251
+▁Nice	-10.0255
+▁teach	-10.0261
+▁Har	-10.0274
+▁cream	-10.0285
+zi	-10.0291
+vin	-10.0295
+▁tools	-10.0319
+▁ground	-10.0324
+▁humans	-10.0343
+je	-10.0355
+ending	-10.036
+minute	-10.0361
+▁target	-10.0363
+IN	-10.0363
+▁quick	-10.039
+dia	-10.0391
+▁condition	-10.0403
+▁visual	-10.0403
+nch	-10.0411
+▁oil	-10.0412
+book	-10.0426
+coming	-10.0432
+mark	-10.044
+▁fashion	-10.0442
+▁prefer	-10.0442
+▁attention	-10.0442
+▁modern	-10.0442
+▁attract	-10.0445
+▁Ty	-10.0455
+▁six	-10.0456
+▁calling	-10.0462
+▁involved	-10.0463
+▁animal	-10.0465
+light	-10.0479
+▁approach	-10.0482
+▁measure	-10.0482
+▁Br	-10.0488
+▁fat	-10.0488
+force	-10.0495
+gent	-10.0498
+▁Lu	-10.0508
+▁topic	-10.0515
+ain	-10.0515
+▁themselves	-10.0522
+▁Everyone	-10.0523
+▁forgot	-10.0524
+▁techniques	-10.0541
+how	-10.0542
+▁Ge	-10.0549
+▁ride	-10.055
+▁president	-10.0558
+▁goals	-10.056
+▁gift	-10.0562
+▁dangerous	-10.0563
+▁concern	-10.0572
+▁fucking	-10.0591
+▁reach	-10.0602
+▁daily	-10.0603
+▁bomb	-10.0603
+▁Sc	-10.0603
+▁cute	-10.0605
+▁1:	-10.0619
+▁page	-10.0625
+▁includes	-10.0631
+col	-10.0632
+lon	-10.0633
+▁political	-10.0643
+▁correct	-10.0643
+▁certain	-10.067
+▁employee	-10.0685
+▁absolutely	-10.0687
+▁ban	-10.0693
+▁implement	-10.0694
+lor	-10.071
+▁Ni	-10.0723
+▁strategies	-10.0725
+▁Given	-10.0725
+▁bread	-10.0728
+▁spot	-10.0731
+▁specifically	-10.0733
+ook	-10.0743
+▁dear	-10.0749
+Yes	-10.0754
+▁struggling	-10.0766
+▁middle	-10.0766
+▁Bra	-10.0766
+▁sweetie	-10.0771
+▁link	-10.0773
+▁realize	-10.078
+▁Cor	-10.0781
+back	-10.0788
+▁Introduction	-10.0807
+▁whether	-10.0807
+of	-10.0832
+ney	-10.0844
+▁confident	-10.0848
+▁security	-10.0848
+▁pop	-10.0852
+▁passage	-10.0853
+based	-10.0864
+▁truth	-10.089
+▁curious	-10.089
+▁Was	-10.0893
+hop	-10.0895
+▁scary	-10.0902
+▁Com	-10.0915
+▁mod	-10.0918
+▁Fe	-10.0919
+▁understandable	-10.0925
+▁shopping	-10.0932
+▁express	-10.0934
+▁tr	-10.0961
+bri	-10.0965
+▁five	-10.0966
+▁ticket	-10.0973
+▁photo	-10.0975
+▁training	-10.0978
+▁camera	-10.0992
+▁Hello	-10.0994
+parent	-10.0994
+▁classes	-10.1002
+▁evidence	-10.1015
+craft	-10.1015
+▁amount	-10.1015
+▁Plus	-10.1016
+▁killed	-10.1018
+▁bike	-10.1023
+▁Mon	-10.1037
+▁prove	-10.105
+▁German	-10.1057
+▁attend	-10.1073
+▁Cha	-10.1087
+▁Definitely	-10.11
+▁climate	-10.11
+▁third	-10.11
+▁products	-10.1118
+▁finished	-10.1138
+▁cheese	-10.1139
+▁trick	-10.1144
+stan	-10.1163
+nny	-10.1164
+▁complain	-10.1185
+▁prevent	-10.1185
+▁boring	-10.1186
+▁student	-10.1189
+▁Two	-10.1193
+▁shape	-10.12
+▁pr	-10.1205
+▁turned	-10.1229
+ties	-10.1233
+▁Tra	-10.124
+▁changes	-10.1249
+▁changed	-10.1255
+▁Get	-10.1269
+️	-10.1271
+▁degree	-10.1274
+rd	-10.1276
+▁vegan	-10.1276
+▁stress	-10.1282
+EN	-10.1286
+▁organ	-10.1295
+▁site	-10.1298
+▁file	-10.1299
+▁analysis	-10.1314
+▁spread	-10.1315
+ified	-10.1322
+ari	-10.1347
+▁North	-10.136
+▁moved	-10.1371
+anna	-10.1372
+pur	-10.1377
+Step	-10.138
+▁reasons	-10.1385
+▁keeping	-10.1385
+▁Tr	-10.1392
+▁Congratulations	-10.1401
+▁himself	-10.1401
+▁recommendations	-10.1402
+▁snow	-10.1402
+ox	-10.1406
+▁gr	-10.1428
+▁vegetables	-10.1434
+▁drop	-10.1445
+▁variety	-10.1445
+▁companies	-10.1445
+▁foot	-10.1446
+▁commit	-10.1448
+▁considered	-10.1448
+▁whi	-10.1451
+▁Most	-10.1454
+▁Bri	-10.1455
+▁legal	-10.1464
+▁drugs	-10.1471
+▁carry	-10.1475
+you	-10.1485
+▁enjoyed	-10.1485
+▁States	-10.1487
+Do	-10.1487
+▁description	-10.1489
+sexual	-10.1504
+▁Wi	-10.1509
+▁soul	-10.1509
+▁Ya	-10.1516
+▁connect	-10.1528
+ization	-10.1536
+▁brain	-10.1536
+▁Per	-10.1565
+▁enter	-10.1574
+▁promote	-10.1578
+▁services	-10.161
+▁regular	-10.1623
+▁truly	-10.1623
+▁Corona	-10.1623
+▁input	-10.1625
+ker	-10.1644
+▁exam	-10.1647
+▁incorporate	-10.1661
+▁internet	-10.1668
+▁cash	-10.1671
+▁ignor	-10.1672
+▁sitting	-10.1673
+▁Col	-10.1689
+▁ill	-10.169
+▁Chi	-10.1691
+▁issues	-10.1695
+nie	-10.1701
+▁realized	-10.1706
+related	-10.1708
+▁Should	-10.1713
+▁Her	-10.1715
+▁towards	-10.1722
+▁blow	-10.1733
+▁mention	-10.1737
+▁seriously	-10.1737
+▁San	-10.174
+ES	-10.1749
+▁excellent	-10.1758
+▁financial	-10.1758
+▁football	-10.1758
+▁frustrating	-10.1758
+▁creative	-10.1758
+▁feature	-10.1765
+▁buying	-10.177
+nut	-10.1788
+▁mu	-10.1797
+gri	-10.1801
+▁within	-10.1804
+▁wanna	-10.1804
+40	-10.1805
+▁rob	-10.1811
+You	-10.1812
+▁Jack	-10.1816
+▁manage	-10.1817
+▁waste	-10.1846
+▁journey	-10.1849
+▁Anything	-10.1849
+▁task	-10.1849
+▁Even	-10.185
+air	-10.1853
+▁girls	-10.1854
+mail	-10.1855
+▁sea	-10.1878
+14	-10.1887
+atic	-10.189
+▁straight	-10.1895
+▁Donald	-10.1895
+▁club	-10.1897
+▁earth	-10.1898
+heat	-10.1898
+▁influence	-10.1909
+▁painting	-10.1933
+▁impressive	-10.1941
+▁summarize	-10.1941
+▁protest	-10.1941
+▁driver	-10.1944
+16	-10.1946
+He	-10.1947
+▁currently	-10.1963
+▁happening	-10.1973
+▁detail	-10.1985
+▁conversation	-10.1987
+▁Europe	-10.1995
+leader	-10.2
+▁farm	-10.2007
+ush	-10.201
+▁similar	-10.203
+▁organization	-10.2034
+za	-10.2038
+▁grab	-10.2052
+▁harm	-10.2071
+▁paying	-10.2088
+!!!	-10.209
+▁budget	-10.2127
+▁Right	-10.2133
+▁compare	-10.2134
+▁wide	-10.214
+iff	-10.2146
+▁players	-10.2157
+▁Joe	-10.2165
+▁fighting	-10.2166
+hal	-10.2172
+▁accomplish	-10.2174
+▁Fo	-10.2177
+▁complete	-10.2208
+tory	-10.2213
+^	-10.2222
+▁Google	-10.2222
+▁Obama	-10.2222
+ED	-10.2226
+ned	-10.226
+▁origin	-10.2268
+▁bird	-10.227
+instruction	-10.2276
+ung	-10.2282
+TH	-10.2297
+ou	-10.2307
+▁challenge	-10.2308
+▁comment	-10.2312
+▁negative	-10.2317
+▁sentence	-10.2318
+▁wild	-10.232
+▁quickly	-10.2334
+▁alive	-10.2335
+▁broke	-10.2353
+▁Actually	-10.2365
+▁French	-10.2365
+▁sauce	-10.2365
+▁stories	-10.2365
+▁novel	-10.2367
+▁flu	-10.2373
+▁trade	-10.2395
+▁society	-10.2414
+▁South	-10.2416
+▁Alex	-10.2416
+▁partner	-10.2422
+▁somewhere	-10.2425
+phone	-10.2432
+standing	-10.2455
+▁environment	-10.2455
+▁begin	-10.2461
+▁Black	-10.2464
+▁Min	-10.2464
+▁AI	-10.2469
+Ti	-10.2489
+haha	-10.2508
+▁patient	-10.2511
+▁medical	-10.2512
+▁jump	-10.2512
+ense	-10.2516
+▁auto	-10.2527
+▁gold	-10.2541
+▁cheat	-10.2557
+▁increase	-10.256
+▁location	-10.256
+▁brought	-10.256
+▁pray	-10.2563
+pec	-10.2563
+▁Ste	-10.2573
+▁tax	-10.2575
+bin	-10.2577
+ING	-10.2577
+ants	-10.2588
+▁struggle	-10.2609
+cc	-10.2621
+ong	-10.2633
+▁Wal	-10.2636
+car	-10.2638
+OP	-10.264
+▁Si	-10.265
+▁Develop	-10.2659
+▁purchase	-10.2659
+ute	-10.2661
+inary	-10.2667
+del	-10.267
+▁setting	-10.2673
+From	-10.2676
+▁def	-10.2686
+▁tea	-10.2697
+gram	-10.27
+▁grocery	-10.2709
+▁library	-10.2709
+▁routine	-10.2709
+▁Tu	-10.2717
+▁East	-10.2718
+▁created	-10.2721
+ica	-10.2735
+▁para	-10.2745
+▁concert	-10.2759
+▁pleasure	-10.2759
+▁vacation	-10.2759
+▁church	-10.2759
+▁characters	-10.2762
+▁option	-10.2764
+▁shoot	-10.277
+list	-10.2772
+▁considering	-10.2777
+▁Time	-10.2788
+ery	-10.2803
+500	-10.2808
+▁contribute	-10.2809
+▁speed	-10.281
+▁invest	-10.281
+▁temp	-10.2814
+ache	-10.2852
+▁allowed	-10.2855
+▁Try	-10.2858
+▁basic	-10.2858
+▁2020	-10.2862
+▁management	-10.2863
+▁practices	-10.2867
+▁wind	-10.2871
+▁cow	-10.2882
+gro	-10.2883
+▁notice	-10.2892
+verse	-10.2906
+▁alternative	-10.291
+▁providing	-10.291
+▁boyfriend	-10.291
+▁series	-10.2912
+▁movement	-10.2936
+▁effectively	-10.2938
+fully	-10.2942
+▁properly	-10.2961
+▁maintain	-10.2963
+▁acc	-10.2984
+▁tall	-10.2985
+▁claim	-10.2987
+az	-10.3001
+60	-10.3003
+▁brings	-10.3005
+down	-10.3005
+▁easier	-10.3012
+▁regard	-10.3013
+▁court	-10.3013
+▁anywhere	-10.3018
+▁En	-10.3033
+▁Pri	-10.3036
+ograph	-10.3037
+25	-10.3041
+off	-10.3056
+😂	-10.3064
+▁fell	-10.3069
+hole	-10.308
+LE	-10.3082
+ino	-10.3094
+yes	-10.3099
+▁customers	-10.3109
+stream	-10.3116
+▁balance	-10.3116
+▁White	-10.3116
+race	-10.3119
+lift	-10.3128
+gree	-10.3142
+▁schedule	-10.3168
+▁solution	-10.3168
+▁period	-10.3168
+▁salad	-10.3168
+▁magic	-10.317
+latest	-10.3175
+Z	-10.318
+▁camp	-10.3185
+▁Love	-10.3191
+▁songs	-10.3203
+▁Pi	-10.3219
+▁challenging	-10.322
+▁literally	-10.322
+▁reduce	-10.3239
+▁custom	-10.3253
+▁pla	-10.3254
+▁Christian	-10.3273
+▁feet	-10.3275
+▁Take	-10.3281
+▁awful	-10.3283
+▁King	-10.3298
+▁City	-10.332
+▁practicing	-10.3326
+▁significance	-10.3326
+▁credit	-10.3326
+▁Earth	-10.3326
+▁lesson	-10.3332
+▁homework	-10.3332
+▁drug	-10.3348
+▁Ko	-10.3359
+late	-10.3372
+▁prop	-10.3373
+▁collect	-10.3373
+▁fantastic	-10.3379
+pit	-10.338
+▁missing	-10.3396
+▁updated	-10.3407
+ati	-10.3415
+▁essential	-10.3432
+▁trend	-10.3447
+▁Dar	-10.3452
+ik	-10.3467
+▁diet	-10.3468
+▁according	-10.3486
+▁earlier	-10.3486
+▁necessary	-10.3486
+▁Accord	-10.3486
+▁embarrass	-10.3486
+▁beauty	-10.3486
+▁mouth	-10.3486
+▁doubt	-10.3487
+▁hide	-10.3491
+free	-10.3498
+ale	-10.35
+▁India	-10.3503
+▁rules	-10.3512
+MA	-10.3513
+▁cro	-10.3523
+▁tend	-10.3526
+▁attitude	-10.354
+▁floor	-10.354
+▁bunch	-10.3542
+▁discuss	-10.355
+▁nor	-10.3588
+▁dev	-10.359
+▁holiday	-10.3595
+▁suffer	-10.3595
+▁damn	-10.3595
+▁dumb	-10.3595
+▁hunt	-10.3598
+▁bright	-10.3598
+▁designed	-10.3601
+load	-10.3613
+▁flat	-10.3615
+▁acting	-10.362
+vy	-10.3631
+▁Aww	-10.3649
+▁THE	-10.3658
+ather	-10.3698
+▁poison	-10.3704
+▁leaving	-10.3704
+▁excuse	-10.3705
+▁election	-10.3706
+Re	-10.3707
+▁Make	-10.3714
+▁Read	-10.3719
+NA	-10.3722
+▁below	-10.3726
+corn	-10.3728
+over	-10.3731
+▁drinking	-10.3735
+▁Clinton	-10.3759
+▁global	-10.3759
+▁peaceful	-10.3759
+▁turnip	-10.3763
+SE	-10.3793
+▁conf	-10.3793
+▁Indian	-10.3804
+ALL	-10.3813
+▁officer	-10.3813
+▁represent	-10.3815
+▁biggest	-10.3815
+▁heavy	-10.3815
+▁region	-10.3815
+ua	-10.3837
+hose	-10.3839
+▁talent	-10.384
+▁Ken	-10.3843
+▁butter	-10.385
+gh	-10.3864
+▁World	-10.3871
+▁Covid	-10.3871
+▁loud	-10.3873
+▁rough	-10.3873
+illa	-10.3882
+80	-10.3883
+fer	-10.3886
+fic	-10.3889
+▁hath	-10.3897
+▁staying	-10.3914
+▁hobbies	-10.3927
+▁client	-10.3927
+▁stock	-10.3927
+▁outdoor	-10.3927
+▁hungry	-10.3928
+▁comments	-10.3938
+▁Art	-10.3951
+▁quit	-10.3956
+▁gotten	-10.3963
+▁universe	-10.3984
+▁truck	-10.3985
+▁device	-10.3985
+▁blog	-10.3992
+▁winter	-10.3999
+▁growing	-10.4007
+▁soup	-10.4007
+▁paint	-10.4009
+21	-10.4039
+▁narrative	-10.404
+▁succeed	-10.404
+▁Provide	-10.404
+▁element	-10.404
+▁forgive	-10.4041
+▁lived	-10.4051
+▁Once	-10.4054
+▁boat	-10.4069
+▁stressed	-10.4081
+stone	-10.4086
+▁sent	-10.4089
+rated	-10.409
+▁audience	-10.4097
+▁concept	-10.4097
+▁speech	-10.4097
+▁request	-10.4098
+▁emotion	-10.4098
+▁besides	-10.4102
+place	-10.4105
+▁owner	-10.4141
+▁mostly	-10.4146
+Christ	-10.4152
+▁drunk	-10.4155
+▁Work	-10.4155
+▁results	-10.4157
+▁lady	-10.4159
+Co	-10.4162
+blo	-10.4162
+▁Ye	-10.4175
+▁total	-10.4185
+▁Dear	-10.4196
+play	-10.42
+▁seat	-10.4212
+▁jealous	-10.4213
+▁interview	-10.4213
+room	-10.4223
+right	-10.4247
+▁effects	-10.4265
+ko	-10.427
+▁platform	-10.4271
+ought	-10.4273
+▁chat	-10.4276
+style	-10.4281
+burg	-10.4282
+hin	-10.429
+▁surprised	-10.4297
+▁Bi	-10.4315
+▁historical	-10.4329
+▁knowledge	-10.4329
+ribu	-10.433
+▁quiet	-10.4331
+▁apartment	-10.4333
+AM	-10.4334
+▁speaking	-10.4345
+yl	-10.4347
+▁chill	-10.4352
+person	-10.4354
+▁require	-10.4356
+life	-10.4364
+▁University	-10.4388
+▁quarantine	-10.4388
+▁subject	-10.4388
+▁damage	-10.4388
+▁document	-10.4388
+▁relaxing	-10.4392
+▁apart	-10.4393
+zo	-10.4396
+lyn	-10.4396
+▁overall	-10.4406
+▁noble	-10.4418
+▁members	-10.442
+IT	-10.4425
+▁Up	-10.4425
+AP	-10.4427
+▁letting	-10.4441
+▁ocean	-10.4447
+▁adult	-10.4448
+▁snack	-10.4448
+scient	-10.4449
+▁brown	-10.4454
+▁blame	-10.446
+▁reality	-10.4465
+▁individuals	-10.4474
+▁sales	-10.4475
+via	-10.448
+house	-10.4485
+▁bat	-10.4493
+▁final	-10.4494
+▁Lord	-10.4509
+▁transportation	-10.451
+▁villager	-10.4513
+▁gain	-10.4515
+▁Same	-10.4523
+▁older	-10.4529
+▁dating	-10.4531
+▁State	-10.4564
+▁emotional	-10.4566
+▁annoying	-10.4567
+▁everywhere	-10.4568
+▁bitch	-10.4576
+▁putt	-10.4581
+In	-10.4588
+▁random	-10.4627
+▁aspect	-10.4627
+▁fli	-10.4636
+▁Mary	-10.4636
+pper	-10.464
+▁manager	-10.4653
+▁string	-10.4654
+▁skill	-10.4662
+▁Saturday	-10.4688
+Muslim	-10.4688
+▁complex	-10.4688
+▁further	-10.4688
+▁shoes	-10.4688
+▁remove	-10.4689
+!"	-10.4693
+▁harder	-10.4695
+13	-10.4699
+ama	-10.4706
+▁update	-10.4717
+hold	-10.4718
+▁Honestly	-10.4749
+▁behavior	-10.4749
+▁responsible	-10.4749
+▁above	-10.4752
+▁perfectly	-10.4753
+▁powerful	-10.4756
+▁station	-10.4762
+▁nobody	-10.4764
+▁Rome	-10.4764
+▁Bro	-10.4776
+lash	-10.4784
+▁Got	-10.4795
+▁Japanese	-10.481
+▁bored	-10.481
+▁mood	-10.4812
+▁gentle	-10.4813
+▁hearing	-10.4817
+▁involve	-10.484
+▁prepared	-10.4841
+▁clearly	-10.4844
+wee	-10.4853
+▁Bill	-10.4867
+▁exploring	-10.4872
+▁respond	-10.4872
+▁destroy	-10.4872
+▁guitar	-10.4872
+▁produce	-10.4872
+▁parasol	-10.4872
+▁policy	-10.4872
+▁adapt	-10.4872
+▁hack	-10.4886
+▁Mc	-10.4888
+div	-10.4889
+▁salt	-10.4923
+▁brief	-10.4934
+▁except	-10.4934
+▁dirty	-10.4934
+▁pressure	-10.4934
+▁neighborhood	-10.4934
+▁More	-10.4934
+▁sharing	-10.4935
+▁section	-10.4936
+▁round	-10.4938
+▁shame	-10.4939
+▁benefit	-10.4949
+▁adding	-10.4955
+▁Say	-10.4964
+▁curse	-10.4987
+▁apply	-10.4988
+▁emb	-10.4994
+▁advance	-10.4996
+▁military	-10.4996
+▁strategy	-10.4996
+▁blind	-10.4996
+▁prison	-10.4997
+Jesus	-10.4999
+▁kick	-10.5001
+▁terms	-10.5001
+duct	-10.5002
+▁Very	-10.5004
+▁mini	-10.503
+▁referr	-10.5041
+year	-10.5046
+▁vegetarian	-10.5059
+▁spirit	-10.5059
+▁dancing	-10.5059
+Gra	-10.5067
+▁yo	-10.507
+▁Roman	-10.5075
+▁Haha	-10.508
+▁dick	-10.5088
+▁prepare	-10.509
+grand	-10.51
+▁cell	-10.5104
+17	-10.5107
+▁spec	-10.5116
+▁becoming	-10.5121
+▁encourage	-10.5122
+▁eventually	-10.5122
+▁engage	-10.5122
+▁faith	-10.5122
+▁finger	-10.5126
+PS	-10.5135
+▁growth	-10.5137
+▁comm	-10.5149
+▁holding	-10.5152
+▁noise	-10.5154
+ification	-10.5183
+▁Twitter	-10.5186
+▁background	-10.5186
+▁perspective	-10.5186
+▁Friday	-10.5186
+▁reddit	-10.5186
+▁expert	-10.5187
+▁threat	-10.5187
+gate	-10.5187
+▁importance	-10.5187
+bel	-10.5197
+▁rice	-10.521
+▁passed	-10.5211
+▁member	-10.5218
+ov	-10.5226
+lap	-10.5243
+▁walked	-10.5243
+▁added	-10.5246
+▁divorce	-10.525
+▁entanglement	-10.525
+▁pandemic	-10.525
+▁confirm	-10.525
+▁cousin	-10.525
+▁adopt	-10.5251
+▁weren	-10.5251
+▁sky	-10.5257
+▁classic	-10.5267
+▁qua	-10.5275
+How	-10.5283
+▁wine	-10.5295
+▁planet	-10.5301
+Russia	-10.5313
+▁Smith	-10.5315
+▁female	-10.5315
+▁religion	-10.5315
+▁1/2	-10.5315
+▁changing	-10.5315
+▁weapon	-10.5315
+▁thousand	-10.5315
+▁appeal	-10.5315
+▁dude	-10.5317
+▁assist	-10.5318
+▁Cat	-10.5321
+non	-10.5327
+▁factors	-10.5352
+▁apple	-10.5375
+▁judge	-10.538
+▁adventure	-10.538
+▁mobile	-10.538
+▁disease	-10.538
+▁shock	-10.538
+▁corona	-10.5383
+▁bake	-10.5384
+▁poet	-10.5385
+▁obviously	-10.5386
+cla	-10.5401
+▁Nope	-10.5408
+ena	-10.5415
+▁useful	-10.5417
+direct	-10.543
+▁Kar	-10.5435
+AD	-10.5436
+▁announce	-10.5445
+▁anxiety	-10.5445
+▁paternity	-10.5445
+▁promotion	-10.5445
+▁traffic	-10.5445
+▁wrote	-10.5445
+▁wedding	-10.5445
+fri	-10.5446
+lthough	-10.5446
+▁master	-10.5447
+▁weak	-10.5455
+▁among	-10.5456
+▁broken	-10.5462
+▁statement	-10.5468
+▁offering	-10.5505
+▁context	-10.5511
+▁kitchen	-10.5511
+▁volunteer	-10.5511
+▁Doctor	-10.5511
+▁swear	-10.5511
+▁Japan	-10.5511
+▁spell	-10.5512
+▁yell	-10.5516
+▁solve	-10.5516
+▁Today	-10.5524
+▁tear	-10.5537
+‘	-10.5577
+▁bottle	-10.5577
+▁economy	-10.5577
+▁private	-10.5577
+▁Hillary	-10.5577
+▁album	-10.5577
+▁admit	-10.5577
+▁electric	-10.5577
+▁appear	-10.558
+▁rule	-10.5601
+▁invite	-10.5608
+▁sup	-10.561
+▁candy	-10.5611
+poli	-10.5614
+▁ruin	-10.5624
+▁$1	-10.5636
+▁```	-10.5644
+▁argument	-10.5644
+▁easily	-10.5644
+▁official	-10.5644
+▁Title	-10.5644
+▁fool	-10.5646
+▁collection	-10.5651
+▁spent	-10.5656
+▁cancel	-10.5665
+▁sale	-10.5671
+lynn	-10.5679
+▁male	-10.5683
+▁host	-10.5686
+US	-10.5687
+▁killing	-10.5688
+uch	-10.5694
+▁critic	-10.571
+▁temperature	-10.5711
+▁smoke	-10.5712
+▁winning	-10.5714
+▁Abou	-10.5716
+bug	-10.575
+▁impressed	-10.5778
+▁School	-10.5778
+▁crack	-10.5779
+pho	-10.5781
+▁stopped	-10.5785
+▁pee	-10.5801
+▁sight	-10.5839
+▁Spanish	-10.5846
+▁dollars	-10.5847
+▁feedback	-10.5847
+▁saving	-10.5847
+▁highlight	-10.585
+▁Count	-10.5852
+▁items	-10.5862
+▁disco	-10.5871
+imply	-10.588
+▁caused	-10.5884
+▁engineer	-10.5887
+▁iron	-10.5892
+▁basketball	-10.5915
+▁breakfast	-10.5915
+▁communication	-10.5915
+▁pregnant	-10.5915
+▁surround	-10.5915
+▁beginning	-10.5916
+▁hiking	-10.5917
+▁Sal	-10.5947
+▁Gi	-10.5974
+▁explanation	-10.5984
+▁responsibility	-10.5984
+▁unfortunate	-10.5984
+▁itself	-10.5984
+▁predict	-10.5984
+▁multi	-10.5985
+dig	-10.6018
+▁influenced	-10.6032
+▁Comp	-10.6041
+▁plane	-10.6048
+frustrated	-10.6053
+▁criminal	-10.6053
+▁knock	-10.6053
+▁nurse	-10.6053
+▁While	-10.6054
+▁smile	-10.6055
+▁Keep	-10.6056
+▁joy	-10.6058
+yah	-10.6072
+We	-10.6077
+▁stealing	-10.6092
+NY	-10.6105
+Awesome	-10.6123
+efficient	-10.6123
+▁ridiculous	-10.6123
+▁insurance	-10.6123
+▁heaven	-10.6123
+▁YOU	-10.6123
+▁idiot	-10.6123
+fro	-10.6124
+▁active	-10.6125
+▁wake	-10.6126
+▁kiss	-10.6128
+residential	-10.6131
+▁Ben	-10.6132
+▁faced	-10.6158
+bil	-10.6158
+▁flowers	-10.6159
+▁workout	-10.6162
+–	-10.6194
+▁citizen	-10.6194
+▁sustainable	-10.6194
+stitution	-10.6194
+▁network	-10.6194
+▁pepper	-10.6194
+▁kept	-10.6195
+▁Angel	-10.6198
+▁gene	-10.6201
+▁soft	-10.6221
+▁surprise	-10.6232
+▁Being	-10.6246
+▁nose	-10.6246
+▁Men	-10.6261
+ept	-10.6263
+▁Sunday	-10.6265
+▁ancient	-10.6265
+▁Italian	-10.6265
+break	-10.6266
+▁treatment	-10.6277
+▁engine	-10.6294
+head	-10.6296
+hel	-10.633
+`	-10.6336
+▁digital	-10.6336
+▁studies	-10.6336
+▁museum	-10.6336
+▁built	-10.6336
+▁frisk	-10.6336
+▁ensure	-10.6336
+▁communities	-10.6336
+▁gym	-10.6337
+▁swim	-10.6337
+rc	-10.6393
+▁dry	-10.6394
+▁festival	-10.6408
+▁instrument	-10.6408
+▁basically	-10.641
+▁House	-10.6411
+▁pound	-10.6412
+▁shut	-10.6413
+face	-10.6421
+even	-10.6425
+▁selling	-10.6428
+▁lying	-10.6428
+▁Home	-10.6436
+▁bless	-10.6444
+uh	-10.6452
+▁hum	-10.6455
+▁comfort	-10.6477
+▁Paris	-10.6479
+▁High	-10.6479
+▁California	-10.648
+▁apologize	-10.648
+▁bunker	-10.648
+▁practical	-10.648
+▁confused	-10.6481
+▁escape	-10.6481
+▁teargas	-10.6481
+▁uncle	-10.6481
+▁demand	-10.6481
+▁Russian	-10.6482
+▁scar	-10.6494
+▁visiting	-10.6498
+▁score	-10.6502
+AL	-10.6505
+▁watched	-10.6517
+▁Lol	-10.6529
+power	-10.655
+▁episode	-10.6553
+▁folklore	-10.6553
+▁contract	-10.6554
+▁losing	-10.6556
+▁punch	-10.6561
+▁Park	-10.6561
+▁safety	-10.6564
+▁rent	-10.6569
+RO	-10.6579
+CK	-10.658
+▁refer	-10.6583
+▁Too	-10.6589
+▁defend	-10.6595
+ette	-10.6619
+▁expected	-10.6622
+▁algorithm	-10.6627
+▁became	-10.6627
+▁Since	-10.6628
+▁baking	-10.6629
+NE	-10.6633
+▁connection	-10.6636
+▁bigger	-10.6642
+99	-10.6647
+▁decor	-10.6654
+elli	-10.6659
+▁Only	-10.666
+23	-10.6661
+▁deliver	-10.6683
+▁mid	-10.6683
+▁Lea	-10.6696
+▁Probably	-10.6701
+▁celebrate	-10.6701
+▁violence	-10.6701
+▁causing	-10.6701
+▁civil	-10.6701
+▁swimming	-10.6701
+▁cheap	-10.6703
+ough	-10.6716
+▁farmer	-10.6716
+▁honestly	-10.6738
+▁sandwich	-10.6776
+▁virtual	-10.6776
+▁National	-10.6776
+▁screen	-10.6776
+BC	-10.6781
+▁flight	-10.6783
+▁writer	-10.679
+▁center	-10.679
+▁Mark	-10.6791
+▁loss	-10.6795
+angle	-10.6798
+▁accus	-10.681
+▁reflect	-10.6817
+▁pie	-10.682
+▁radi	-10.6821
+▁glass	-10.6843
+▁memories	-10.6851
+▁crowd	-10.6851
+▁corner	-10.6859
+▁Never	-10.6864
+hood	-10.6871
+▁Stu	-10.6875
+CO	-10.6887
+▁map	-10.6891
+▁suppose	-10.6895
+▁fail	-10.6909
+▁sleeping	-10.6911
+▁memory	-10.6928
+▁effect	-10.6931
+▁included	-10.6939
+▁determined	-10.6948
+LA	-10.6948
+▁disappointed	-10.6949
+▁grace	-10.6955
+▁feed	-10.6962
+90	-10.6974
+▁strength	-10.7003
+▁summary	-10.7003
+▁lonely	-10.7003
+▁multiple	-10.7004
+▁poop	-10.7007
+▁replace	-10.7011
+▁Fun	-10.7017
+▁metal	-10.7021
+ford	-10.7038
+▁held	-10.7043
+▁equal	-10.7045
+▁successfully	-10.705
+▁throughou	-10.7055
+▁Chan	-10.7056
+▁bla	-10.706
+logical	-10.7063
+aught	-10.7072
+▁beyond	-10.708
+▁photograph	-10.708
+▁England	-10.708
+▁lazy	-10.708
+▁TIL	-10.708
+▁habit	-10.708
+▁belong	-10.708
+▁tomato	-10.7081
+▁Gar	-10.7082
+▁Jim	-10.7088
+▁Tri	-10.7099
+factor	-10.7107
+19	-10.712
+▁counter	-10.714
+▁actual	-10.7144
+▁Mexican	-10.7157
+▁epicenter	-10.7157
+▁quote	-10.7157
+▁crash	-10.7157
+▁graphic	-10.7157
+▁twice	-10.7158
+▁contain	-10.716
+▁2016	-10.716
+▁pasta	-10.7205
+▁spin	-10.7235
+▁Python	-10.7236
+▁advantage	-10.7236
+▁population	-10.7236
+▁terrorist	-10.7236
+▁British	-10.7236
+▁install	-10.7236
+▁sugar	-10.7236
+▁motivated	-10.7245
+proof	-10.7253
+▁mass	-10.7262
+▁Uh	-10.7267
+▁required	-10.7279
+PA	-10.7287
+▁sport	-10.7293
+fec	-10.7311
+▁Yo	-10.7313
+▁experiment	-10.7314
+▁psych	-10.7315
+▁grew	-10.7316
+▁passionate	-10.7317
+▁trash	-10.7318
+▁Someone	-10.7318
+iness	-10.7321
+▁billion	-10.7327
+▁singing	-10.7343
+▁Wh	-10.7362
+▁Sir	-10.7377
+▁affected	-10.7394
+▁mountain	-10.7394
+▁survive	-10.7394
+▁beef	-10.7395
+▁entertain	-10.7395
+▁pretend	-10.7396
+▁conduct	-10.7396
+▁interact	-10.7396
+▁environmental	-10.7405
+▁located	-10.7424
+▁Nah	-10.7427
+NO	-10.7436
+▁renewable	-10.7474
+▁average	-10.7474
+ternal	-10.7479
+crib	-10.7488
+▁24	-10.7504
+ez	-10.7512
+▁implemented	-10.7534
+▁concerned	-10.7535
+CH	-10.7536
+▁vari	-10.7551
+▁rape	-10.7552
+▁Lily	-10.7554
+—	-10.7554
+▁creature	-10.7554
+▁Scott	-10.7554
+▁minor	-10.7555
+▁porn	-10.7559
+▁national	-10.7562
+▁pants	-10.7564
+▁suddenly	-10.7567
+▁addict	-10.7571
+▁odd	-10.7575
+▁Mil	-10.7603
+0,000	-10.7618
+&	-10.7625
+▁object	-10.7633
+▁competition	-10.7636
+▁hundred	-10.7636
+▁journal	-10.7636
+▁stomach	-10.7636
+▁victim	-10.7636
+▁property	-10.7636
+▁Also	-10.7637
+▁France	-10.7639
+▁indeed	-10.764
+▁rush	-10.7642
+▁standard	-10.7643
+▁Johnson	-10.7644
+▁lack	-10.7648
+▁fishing	-10.7651
+▁genre	-10.7654
+atter	-10.7662
+▁incorporat	-10.7672
+flower	-10.7677
+RA	-10.7677
+▁seven	-10.7681
+▁shooting	-10.7699
+▁voted	-10.7714
+▁Richard	-10.7718
+▁Unfortunately	-10.7718
+▁collaborat	-10.7718
+▁enhance	-10.7718
+▁ventilator	-10.7718
+▁false	-10.7718
+▁tennis	-10.7718
+▁lawyer	-10.7718
+▁launch	-10.7718
+▁healthcare	-10.772
+▁Using	-10.7726
+▁consume	-10.7726
+▁Show	-10.7736
+point	-10.7743
+▁however	-10.7745
+▁West	-10.7754
+▁grown	-10.7754
+▁Za	-10.7756
+brow	-10.776
+▁talented	-10.7763
+▁payment	-10.7767
+▁mindfulness	-10.78
+▁replied	-10.78
+▁Design	-10.78
+▁battle	-10.7801
+▁Yu	-10.7803
+▁brave	-10.7811
+▁buck	-10.7814
+▁wave	-10.7824
+coin	-10.7844
+▁none	-10.7853
+▁structure	-10.7861
+▁Cap	-10.7872
+▁horror	-10.7884
+▁Something	-10.7884
+ET	-10.7885
+▁assistant	-10.7886
+▁expand	-10.7888
+▁Wait	-10.7901
+clock	-10.791
+lab	-10.7933
+Well	-10.7938
+fish	-10.794
+▁pattern	-10.7968
+▁remain	-10.7968
+▁Three	-10.7968
+▁transfer	-10.7969
+▁Irish	-10.7972
+▁everyday	-10.7979
+▁Jane	-10.7992
+▁Over	-10.7994
+▁cheer	-10.8
+▁locate	-10.802
+▁equ	-10.8029
+▁exact	-10.8048
+▁reduc	-10.8049
+▁Lin	-10.805
+▁baseball	-10.8052
+▁crucial	-10.8052
+▁religious	-10.8052
+▁romantic	-10.8052
+Friend	-10.8052
+▁foreign	-10.8053
+▁prince	-10.8053
+▁drum	-10.8053
+▁critical	-10.8055
+▁robot	-10.8056
+▁lower	-10.8064
+▁musician	-10.8068
+ux	-10.8073
+where	-10.8073
+▁somebody	-10.81
+▁release	-10.8106
+▁Call	-10.8117
+▁Val	-10.8118
+▁cheating	-10.8122
+▁Emily	-10.8138
+▁Republican	-10.8138
+▁Welcome	-10.8138
+▁activity	-10.8138
+▁architecture	-10.8138
+▁clothing	-10.8138
+▁industries	-10.8138
+▁medicine	-10.8138
+▁previous	-10.8138
+▁generation	-10.8138
+▁cloud	-10.8138
+▁wheel	-10.8138
+▁variations	-10.8138
+▁teaching	-10.8138
+▁woke	-10.814
+▁immediately	-10.8141
+▁containment	-10.8141
+▁Grand	-10.8143
+▁ugly	-10.8144
+ella	-10.8146
+▁UK	-10.8148
+ji	-10.8175
+▁blend	-10.818
+lion	-10.8181
+▁individual	-10.8186
+field	-10.8209
+know	-10.8216
+▁London	-10.8224
+▁analyze	-10.8224
+▁cuisine	-10.8224
+▁phrase	-10.8224
+▁policies	-10.8224
+▁afternoon	-10.8224
+▁belief	-10.8224
+▁flour	-10.8225
+70	-10.8229
+oooo	-10.8257
+▁Africa	-10.8299
+▁employ	-10.8309
+▁scratch	-10.8311
+▁channel	-10.8311
+▁plastic	-10.8311
+▁soccer	-10.8311
+▁rewarding	-10.8311
+▁asshole	-10.8312
+▁direction	-10.8312
+▁package	-10.8312
+▁punish	-10.8312
+▁penis	-10.8315
+▁cough	-10.8317
+▁depending	-10.8329
+▁received	-10.8339
+▁Thou	-10.8348
+▁Max	-10.8349
+▁hug	-10.8358
+Net	-10.8365
+▁sudden	-10.8385
+é	-10.8399
+▁Florida	-10.8399
+▁intelligence	-10.8399
+▁marriage	-10.8399
+▁optimiz	-10.8399
+▁managing	-10.8399
+▁bowl	-10.8399
+▁knee	-10.84
+▁tongue	-10.84
+▁trump	-10.84
+▁valid	-10.84
+dependent	-10.84
+▁Super	-10.8408
+▁covered	-10.8413
+▁production	-10.8421
+▁till	-10.8437
+▁copy	-10.8442
+▁coat	-10.8449
+iah	-10.8457
+▁alcohol	-10.8464
+EE	-10.8464
+▁Korea	-10.8479
+▁younger	-10.8482
+\	-10.8487
+appropriate	-10.8487
+▁convince	-10.8487
+▁ourselves	-10.8487
+▁protagonist	-10.8487
+▁honor	-10.8487
+▁Apple	-10.8487
+▁plot	-10.8489
+▁display	-10.849
+▁cities	-10.8491
+▁tight	-10.8491
+▁orange	-10.8492
+▁cri	-10.8496
+▁moon	-10.8496
+ably	-10.8497
+▁hotel	-10.8504
+▁parties	-10.8513
+SH	-10.8522
+RT	-10.8544
+▁European	-10.8561
+volution	-10.8574
+▁bartender	-10.8576
+▁conflict	-10.8576
+▁identify	-10.8576
+▁prompt	-10.8576
+▁David	-10.8577
+▁William	-10.8577
+▁Pretty	-10.8577
+▁bathroom	-10.8577
+▁slice	-10.8578
+▁gather	-10.8579
+▁Char	-10.8643
+take	-10.8646
+▁Bla	-10.865
+GO	-10.866
+liberal	-10.8667
+▁thread	-10.8669
+▁trail	-10.8676
+▁investment	-10.8678
+▁Win	-10.8681
+▁Asian	-10.8685
+LL	-10.8706
+Last	-10.8713
+▁thin	-10.8734
+TS	-10.8737
+▁Yea	-10.8744
+▁dessert	-10.8757
+▁unfair	-10.8757
+▁politics	-10.8757
+▁switch	-10.8758
+▁attempt	-10.8758
+▁double	-10.8758
+▁intense	-10.876
+▁slave	-10.876
+▁hike	-10.8774
+▁passion	-10.8778
+▁discussing	-10.8824
+Yeah	-10.8833
+▁smooth	-10.8838
+▁birth	-10.8848
+▁breast	-10.8849
+▁engaging	-10.8849
+▁typical	-10.8849
+▁Street	-10.8849
+▁intention	-10.8849
+▁homemade	-10.8849
+▁medium	-10.885
+▁prank	-10.885
+▁marry	-10.8852
+▁chair	-10.8852
+▁addition	-10.8857
+▁south	-10.8866
+▁actor	-10.8872
+▁generate	-10.8888
+▁bath	-10.8894
+▁USA	-10.8894
+▁camping	-10.8918
+▁Use	-10.8937
+▁disagree	-10.8942
+▁complicated	-10.8942
+▁garlic	-10.8942
+▁meditation	-10.8942
+▁freak	-10.8942
+▁James	-10.8944
+▁Tim	-10.8945
+rose	-10.8951
+▁settle	-10.8956
+▁Paul	-10.8956
+▁cancer	-10.8965
+▁failed	-10.8979
+▁plate	-10.8985
+▁Won	-10.8992
+▁tag	-10.9019
+▁cul	-10.902
+▁wh	-10.9024
+▁1000	-10.9033
+▁Mexico	-10.9035
+▁consequences	-10.9035
+▁knife	-10.9035
+▁overwhelming	-10.9035
+▁organized	-10.9037
+▁gross	-10.9038
+▁jail	-10.9038
+▁species	-10.9038
+▁babe	-10.9039
+▁huh	-10.9044
+view	-10.9046
+▁Peter	-10.9047
+▁Mike	-10.9051
+conserv	-10.9065
+▁stage	-10.9068
+▁flag	-10.9075
+▁unless	-10.9077
+▁Texas	-10.913
+▁candidate	-10.913
+▁clarify	-10.913
+▁ghost	-10.913
+▁tattoo	-10.913
+▁drama	-10.913
+▁gender	-10.9131
+▁capture	-10.9133
+▁Del	-10.9135
+income	-10.914
+▁personality	-10.9149
+▁Act	-10.917
+▁dying	-10.9177
+water	-10.9187
+▁edge	-10.92
+▁differ	-10.9221
+▁Facebook	-10.9225
+▁international	-10.9225
+▁overwhelmed	-10.9225
+▁College	-10.9225
+▁informative	-10.9225
+▁detect	-10.9225
+▁debt	-10.9226
+▁shout	-10.9227
+▁Canada	-10.9229
+▁shake	-10.9232
+▁constantly	-10.9232
+▁Things	-10.9232
+▁delivery	-10.9249
+grad	-10.9255
+▁ruined	-10.9269
+ility	-10.9282
+▁faster	-10.9294
+Your	-10.9305
+MO	-10.9315
+▁surgery	-10.9321
+▁technical	-10.9321
+▁pogrom	-10.9321
+▁vivid	-10.9321
+▁retail	-10.9321
+▁Besides	-10.9321
+▁commun	-10.9323
+▁transform	-10.9323
+▁alien	-10.9325
+▁reaction	-10.9326
+▁extremely	-10.9331
+▁glasses	-10.9331
+▁spice	-10.9332
+▁relate	-10.9343
+leading	-10.9356
+▁released	-10.9357
+▁Many	-10.9405
+▁Shi	-10.941
+▁result	-10.9414
+Instruction	-10.9418
+▁calculate	-10.9418
+▁favourite	-10.9418
+▁recognize	-10.9418
+▁square	-10.9418
+conscious	-10.9418
+▁April	-10.9418
+▁smoking	-10.9418
+▁utiliz	-10.9418
+▁screw	-10.9419
+▁strip	-10.942
+▁yoga	-10.942
+▁construction	-10.9421
+hand	-10.9421
+▁guest	-10.943
+▁tradition	-10.9465
+bound	-10.9477
+▁painful	-10.9505
+▁exhaust	-10.9516
+▁fictional	-10.9516
+▁guilty	-10.9516
+▁identity	-10.9516
+▁century	-10.9516
+▁assume	-10.9517
+▁adjust	-10.9517
+▁childhood	-10.9517
+▁oven	-10.9518
+▁fellow	-10.952
+▁quant	-10.952
+▁accidentally	-10.952
+▁pool	-10.9523
+▁fitness	-10.9523
+▁neck	-10.9527
+▁Rock	-10.9537
+▁Cook	-10.9541
+▁grant	-10.9545
+OW	-10.9548
+▁etc	-10.9565
+▁whenever	-10.9578
+▁symbol	-10.9614
+migrant	-10.9615
+▁improving	-10.9615
+▁Marketing	-10.9615
+▁desert	-10.9615
+▁desire	-10.9616
+▁charity	-10.9616
+▁password	-10.9618
+▁Anyone	-10.9618
+▁Italy	-10.9627
+▁African	-10.9629
+▁abuse	-10.9636
+headed	-10.9645
+▁pun	-10.9669
+▁combination	-10.9715
+▁uncomfortable	-10.9715
+▁borrow	-10.9715
+▁communicate	-10.9715
+▁opposite	-10.9716
+profit	-10.9716
+▁hiding	-10.9717
+▁decade	-10.9717
+#	-10.9722
+▁wrap	-10.9725
+jack	-10.9732
+dress	-10.974
+▁TO	-10.9769
+▁rise	-10.9772
+▁receive	-10.9784
+▁Wil	-10.9799
+▁format	-10.9814
+▁statistics	-10.9816
+▁Generate	-10.9816
+▁Question	-10.9816
+▁Bernie	-10.9816
+▁micro	-10.9816
+▁evil	-10.9818
+▁coast	-10.9819
+OL	-10.9834
+▁singer	-10.9845
+▁Bob	-10.9852
+▁Both	-10.9879
+▁disappoint	-10.9888
+▁Big	-10.9892
+▁Michael	-10.9918
+▁offensive	-10.9918
+▁regret	-10.9918
+▁hobby	-10.9918
+▁relative	-10.9918
+▁staff	-10.9926
+quart	-10.995
+▁Year	-10.9955
+just	-10.998
+▁Apparently	-11.0021
+▁Remember	-11.0021
+▁license	-11.0021
+▁terrific	-11.0021
+▁publish	-11.0021
+▁honour	-11.0021
+▁capital	-11.0021
+▁freedom	-11.0021
+▁error	-11.0022
+▁spare	-11.0027
+▁Think	-11.0027
+▁presentation	-11.0036
+▁arrested	-11.0043
+rust	-11.0043
+45	-11.0059
+▁despite	-11.0125
+▁fantasy	-11.0125
+▁opportunities	-11.0125
+▁Girl	-11.0125
+▁remote	-11.0125
+▁Social	-11.0126
+▁Health	-11.0126
+▁snake	-11.0126
+▁piano	-11.0126
+▁mystery	-11.0126
+▁Damn	-11.0126
+distance	-11.0126
+VER	-11.0131
+▁thus	-11.0149
+▁lake	-11.0153
+▁crap	-11.0153
+▁slip	-11.0162
+▁poll	-11.0177
+UN	-11.0219
+talk	-11.0229
+▁descriptive	-11.0231
+▁families	-11.0231
+▁medication	-11.0231
+▁hidden	-11.0231
+▁promoting	-11.0231
+▁soldier	-11.0231
+▁teeth	-11.0231
+▁refuse	-11.0231
+▁royal	-11.0231
+▁roast	-11.0231
+▁dropped	-11.0233
+▁True	-11.0233
+▁kidding	-11.0235
+▁Stop	-11.0245
+▁grill	-11.0246
+▁taxes	-11.0248
+▁Out	-11.0265
+leigh	-11.0299
+posed	-11.0312
+▁Maria	-11.0313
+▁Spi	-11.0331
+▁transport	-11.0336
+▁horrible	-11.0337
+▁motivation	-11.0337
+▁sensory	-11.0337
+▁monitor	-11.0337
+▁herself	-11.0337
+▁Hitler	-11.0337
+▁Trust	-11.034
+▁Turn	-11.0342
+▁stressful	-11.0342
+▁nail	-11.0356
+fold	-11.0356
+▁Post	-11.0371
+▁incl	-11.042
+worker	-11.042
+friend	-11.0444
+▁Explain	-11.0445
+▁encounter	-11.0445
+▁George	-11.0445
+▁steak	-11.0445
+▁convert	-11.0445
+▁firm	-11.0446
+▁empty	-11.0446
+▁directly	-11.0448
+Why	-11.0468
+▁vegetable	-11.0473
+▁nation	-11.0492
+▁dedicated	-11.0553
+▁elaborate	-11.0553
+▁possibly	-11.0553
+▁babies	-11.0553
+▁operation	-11.0553
+▁loving	-11.0553
+▁arrive	-11.0554
+▁grandma	-11.0554
+▁priest	-11.0556
+▁programming	-11.0556
+▁resist	-11.0557
+▁plus	-11.0559
+▁potatoes	-11.0559
+▁nowhere	-11.0565
+▁classical	-11.058
+▁earn	-11.0613
+▁rip	-11.0659
+▁Empire	-11.0663
+▁impossible	-11.0663
+▁jewelry	-11.0663
+▁substitute	-11.0663
+▁violent	-11.0663
+▁stood	-11.0663
+▁trial	-11.0667
+▁discussion	-11.0675
+▁taught	-11.0685
+Trump	-11.0709
+▁hip	-11.0774
+length	-11.0774
+▁strike	-11.0774
+▁Monday	-11.0777
+▁queen	-11.0781
+▁Romeo	-11.079
+▁friendship	-11.0792
+▁Fuck	-11.0794
+▁Vol	-11.0798
+UR	-11.0824
+▁ppl	-11.0843
+▁advertising	-11.0887
+▁definition	-11.0887
+▁distancing	-11.0887
+▁mechanic	-11.0887
+inspired	-11.0887
+▁repeat	-11.089
+hhh	-11.0893
+▁react	-11.0937
+▁determine	-11.0968
+▁rank	-11.0976
+catching	-11.0983
+▁$100	-11.0985
+▁register	-11.1
+❤	-11.1
+▁debate	-11.1001
+▁Prince	-11.1002
+▁guard	-11.1005
+▁chain	-11.1009
+awn	-11.1015
+▁Yep	-11.102
+▁duck	-11.1025
+▁yea	-11.1028
+scale	-11.1065
+sensitive	-11.1115
+{	-11.1115
+|	-11.1115
+▁Amazon	-11.1115
+▁equipment	-11.1115
+▁introduce	-11.1115
+▁thorough	-11.1115
+▁Jewish	-11.1115
+▁laptop	-11.1115
+▁confusing	-11.1115
+▁distract	-11.1116
+applied	-11.1116
+▁requirements	-11.1116
+▁diverse	-11.1117
+▁import	-11.1123
+▁River	-11.1127
+▁burger	-11.1136
+▁Greek	-11.1232
+▁vehicle	-11.1232
+▁arguing	-11.1232
+▁separate	-11.1232
+▁navigat	-11.1232
+▁Still	-11.1232
+▁contemporary	-11.1232
+▁hopefully	-11.1237
+▁bio	-11.124
+▁humor	-11.1243
+▁director	-11.1255
+▁dust	-11.1286
+▁establish	-11.1319
+26	-11.1333
+▁admirable	-11.1349
+▁conference	-11.1349
+▁neither	-11.1349
+▁symptoms	-11.1349
+▁teenager	-11.1349
+▁turkey	-11.1349
+▁twitter	-11.1349
+▁Another	-11.1349
+▁DIY	-11.1349
+▁bodies	-11.1349
+▁Before	-11.1349
+▁jerk	-11.135
+▁Help	-11.135
+▁insights	-11.135
+▁Congress	-11.135
+▁repair	-11.1351
+▁roof	-11.1351
+▁river	-11.1356
+▁handling	-11.136
+▁500-	-11.136
+▁comic	-11.1365
+▁Gold	-11.137
+▁blast	-11.1371
+▁Play	-11.1372
+▁Must	-11.1376
+▁Again	-11.1447
+▁counsel	-11.1468
+▁incredibly	-11.1468
+▁stretch	-11.1468
+▁studied	-11.1468
+▁valuable	-11.1468
+▁disaster	-11.1468
+▁protein	-11.1468
+▁Edward	-11.1468
+▁inviting	-11.1469
+Virus	-11.147
+▁desk	-11.1472
+▁command	-11.1472
+▁Office	-11.1474
+▁NOT	-11.1506
+school	-11.1516
+0%	-11.1525
+▁technique	-11.1533
+▁Brea	-11.1549
+▁danger	-11.1588
+▁favor	-11.1588
+▁commercial	-11.1589
+▁confidence	-11.1589
+▁disappear	-11.1589
+▁increasing	-11.1589
+▁proposal	-11.1589
+▁subreddit	-11.1589
+▁chemical	-11.1589
+▁border	-11.1589
+▁climb	-11.1589
+▁loyal	-11.1589
+▁juice	-11.159
+▁wealth	-11.159
+▁Frank	-11.1591
+▁Brown	-11.1591
+▁fuel	-11.1592
+▁preference	-11.1592
+▁beginner	-11.1592
+▁everybody	-11.1596
+▁stolen	-11.1599
+WW	-11.1608
+cular	-11.1609
+aaaa	-11.161
+▁village	-11.1699
+▁awhile	-11.1711
+▁hilarious	-11.1711
+▁mirror	-11.1711
+▁scientific	-11.1711
+▁invented	-11.1711
+▁integrate	-11.1711
+▁presence	-11.1711
+▁reveal	-11.1712
+▁2019	-11.1713
+▁tourist	-11.1713
+▁appearance	-11.1714
+ologist	-11.1714
+▁overview	-11.1717
+▁Cause	-11.1718
+▁define	-11.1725
+▁Old	-11.1726
+$	-11.1732
+ulate	-11.1791
+world	-11.1799
+▁hung	-11.1834
+▁accurate	-11.1834
+▁athlete	-11.1834
+▁destination	-11.1834
+▁evolved	-11.1834
+▁innovative	-11.1834
+▁league	-11.1834
+▁politician	-11.1834
+▁1/4	-11.1834
+▁buddy	-11.1834
+▁podcast	-11.1834
+▁anxious	-11.1834
+▁CEO	-11.1834
+▁otherwise	-11.1834
+shadow	-11.1834
+▁showcase	-11.1834
+▁replies	-11.1834
+▁spider	-11.1834
+▁witness	-11.1834
+▁shelter	-11.1834
+▁pilot	-11.1835
+▁Biden	-11.1836
+▁zero	-11.1837
+▁yellow	-11.1838
+▁wound	-11.1839
+▁Anytime	-11.1839
+▁committed	-11.1839
+▁nearby	-11.1843
+▁gang	-11.1846
+bringing	-11.1846
+▁Water	-11.1849
+▁Report	-11.1851
+horn	-11.1887
+▁chose	-11.1887
+▁panic	-11.1893
+▁bullshit	-11.1959
+▁initiative	-11.1959
+▁landscape	-11.1959
+▁nuclear	-11.1959
+▁occasion	-11.1959
+▁technologies	-11.1959
+▁Chapter	-11.1959
+▁burden	-11.1959
+▁Marcius	-11.1959
+▁compliment	-11.1959
+▁asleep	-11.1959
+finite	-11.196
+unch	-11.1962
+▁originally	-11.1963
+▁array	-11.1963
+▁thick	-11.1967
+▁cutting	-11.1976
+▁Want	-11.2001
+ambi	-11.2066
+▁magazine	-11.2086
+▁muscle	-11.2086
+▁hurry	-11.2086
+▁relief	-11.2086
+▁Police	-11.2086
+▁fraud	-11.2086
+▁profile	-11.2086
+▁Perfect	-11.2086
+▁reliable	-11.2086
+▁naked	-11.2086
+▁secure	-11.2086
+▁therapist	-11.2086
+▁crisp	-11.2087
+▁coworker	-11.2088
+▁attach	-11.2088
+▁loose	-11.2089
+earning	-11.2098
+▁Start	-11.21
+Elect	-11.2114
+▁oppos	-11.2121
+specially	-11.2127
+morrow	-11.2181
+▁choosing	-11.2214
+▁infrastructure	-11.2214
+▁newspaper	-11.2214
+▁relevant	-11.2214
+▁semester	-11.2214
+▁university	-11.2214
+▁monkey	-11.2214
+▁nigga	-11.2214
+▁crisis	-11.2214
+▁crush	-11.2214
+▁advancements	-11.2215
+▁crust	-11.2216
+▁banana	-11.2216
+▁couch	-11.2217
+▁america	-11.2218
+▁Trans	-11.2218
+▁planned	-11.2219
+▁Jeff	-11.2219
+▁Mother	-11.2221
+▁Based	-11.2224
+▁download	-11.2224
+▁moral	-11.2226
+▁capabilit	-11.2333
+▁Washington	-11.2344
+▁bridge	-11.2344
+▁paraphrase	-11.2344
+▁suspense	-11.2344
+▁violin	-11.2344
+▁gluten	-11.2344
+▁admire	-11.2344
+▁enemy	-11.2344
+▁romance	-11.2344
+▁Next	-11.2344
+▁Robert	-11.2345
+;#	-11.2349
+▁Harry	-11.2354
+▁stir	-11.2358
+▁twin	-11.2367
+😭	-11.2475
+▁necessarily	-11.2475
+ville	-11.2477
+▁failure	-11.2477
+▁bedroom	-11.2477
+▁outfit	-11.2478
+▁spray	-11.248
+▁solid	-11.2482
+▁voters	-11.2482
+▁abortion	-11.2609
+▁disorder	-11.2609
+▁exhibit	-11.2609
+▁graduated	-11.2609
+▁grandmother	-11.2609
+▁laundry	-11.2609
+▁possess	-11.2609
+▁Imagine	-11.2609
+▁Little	-11.2609
+▁Women	-11.2609
+▁finance	-11.2609
+▁observ	-11.2609
+▁Israel	-11.2609
+▁scheme	-11.2609
+▁Jones	-11.261
+▁realistic	-11.2615
+▁olive	-11.2617
+hrew	-11.2634
+▁Though	-11.2729
+▁obvious	-11.273
+▁Chicago	-11.2744
+▁Seriously	-11.2744
+▁bottom	-11.2744
+▁circum	-11.2744
+▁draft	-11.2744
+▁suspect	-11.2744
+▁blonde	-11.2744
+▁Science	-11.2744
+▁abilities	-11.2744
+~~	-11.2744
+▁leadership	-11.2744
+▁twist	-11.2744
+▁defense	-11.2745
+depth	-11.2745
+▁banned	-11.2745
+▁objective	-11.2749
+▁extreme	-11.2867
+▁Research	-11.2881
+▁academic	-11.2881
+▁department	-11.2881
+▁gratitude	-11.2881
+▁perhaps	-11.2881
+Tuesday	-11.2881
+▁powder	-11.2881
+▁revenge	-11.2881
+▁ladies	-11.2881
+▁injury	-11.2881
+▁storage	-11.2881
+▁Sweet	-11.2882
+▁peanut	-11.2883
+▁reference	-11.2883
+▁Chris	-11.2887
+▁Find	-11.2888
+tooth	-11.3015
+▁cigarette	-11.302
+▁dedication	-11.302
+▁dragon	-11.302
+▁mushroom	-11.302
+▁necklace	-11.302
+▁survey	-11.302
+▁Canadian	-11.302
+▁YouTube	-11.302
+revised	-11.302
+▁Verse	-11.302
+▁evolution	-11.3023
+▁interactive	-11.3025
+▁zoo	-11.304
+maker	-11.3061
+▁scam	-11.3086
+▁assault	-11.3161
+▁creativity	-11.3161
+▁garbage	-11.3161
+▁incident	-11.3161
+▁primary	-11.3161
+▁brush	-11.3161
+▁Sanders	-11.3161
+▁barely	-11.3161
+▁silly	-11.3162
+▁Queen	-11.3163
+Word	-11.3163
+▁label	-11.3163
+▁massive	-11.3166
+▁spring	-11.3169
+▁litera	-11.3173
+▁Under	-11.3174
+▁Long	-11.3176
+watch	-11.3186
+Enter	-11.3202
+gress	-11.3211
+driven	-11.328
+▁stole	-11.3292
+▁Middle	-11.3303
+▁component	-11.3303
+▁convenient	-11.3303
+▁inspiration	-11.3303
+▁poverty	-11.3303
+▁sweetheart	-11.3303
+▁unhappy	-11.3303
+▁Spring	-11.3303
+▁rural	-11.3303
+▁literature	-11.3304
+▁toxic	-11.3305
+▁symbolism	-11.3305
+▁crown	-11.3305
+▁Charles	-11.3306
+▁LOL	-11.3306
+▁curry	-11.3306
+▁2017	-11.3306
+▁detective	-11.3308
+space	-11.3356
+rtificial	-11.3448
+▁Australian	-11.3448
+▁Instagram	-11.3448
+▁Internet	-11.3448
+▁January	-11.3448
+▁appointment	-11.3448
+▁chopped	-11.3448
+▁competitive	-11.3448
+▁groceries	-11.3448
+▁guidance	-11.3448
+▁stunning	-11.3448
+▁unexpected	-11.3448
+▁serving	-11.3448
+▁worries	-11.3448
+▁broad	-11.3448
+▁jacket	-11.3449
+▁expectations	-11.3449
+▁joking	-11.3449
+▁status	-11.3449
+▁carbon	-11.345
+▁bond	-11.3454
+▁route	-11.3459
+▁smoothie	-11.3467
+▁arrest	-11.3566
+▁Egypt	-11.3595
+▁abandon	-11.3595
+▁dialogue	-11.3595
+▁tutorial	-11.3595
+▁scream	-11.3595
+▁shift	-11.3595
+▁toppings	-11.3596
+▁branch	-11.3596
+▁sushi	-11.3596
+▁therefore	-11.3596
+▁racism	-11.3596
+▁Guess	-11.3596
+▁History	-11.3597
+▁preserv	-11.3598
+▁proposed	-11.3598
+▁voting	-11.36
+▁chef	-11.3603
+▁hook	-11.3608
+▁continu	-11.3628
+build	-11.3693
+▁thrill	-11.3704
+▁congratulations	-11.3745
+▁folks	-11.3745
+▁participate	-11.3745
+▁pocket	-11.3745
+▁professor	-11.3745
+▁suicide	-11.3745
+▁supplies	-11.3745
+▁vaccine	-11.3745
+▁Speak	-11.3745
+▁throat	-11.3745
+▁battery	-11.3745
+▁overcome	-11.3745
+▁strict	-11.3745
+▁occur	-11.3745
+▁League	-11.3746
+▁labor	-11.3747
+▁monster	-11.3747
+▁motion	-11.3748
+▁duty	-11.3748
+▁Ugh	-11.3749
+▁argue	-11.3749
+▁Creat	-11.3826
+muni	-11.3843
+When	-11.3869
+▁Hopefully	-11.3896
+▁Microsoft	-11.3896
+▁happiness	-11.3896
+▁iPhone	-11.3896
+▁nutrition	-11.3896
+▁twenty	-11.3896
+▁possibility	-11.3896
+▁session	-11.3896
+▁Tower	-11.3897
+▁facing	-11.3897
+▁100%	-11.3898
+▁emerg	-11.3898
+▁Talk	-11.39
+▁spoke	-11.3977
+▁December	-11.405
+▁exchange	-11.405
+▁experiencing	-11.405
+▁principle	-11.405
+▁seafood	-11.405
+▁pricing	-11.4051
+▁Guide	-11.4053
+empt	-11.4081
+▁Edit	-11.4107
+Democrats	-11.4178
+▁potato	-11.4198
+~	-11.4206
+▁Michigan	-11.4206
+▁brainstorm	-11.4206
+▁metaphor	-11.4206
+▁nasty	-11.4207
+▁jazz	-11.4207
+▁steel	-11.4207
+▁funeral	-11.4207
+▁discount	-11.4208
+▁surf	-11.4228
+▁broth	-11.4318
+▁Enjoy	-11.4365
+▁Chicken	-11.4365
+▁grief	-11.4365
+▁Disney	-11.4365
+▁General	-11.4365
+▁tribe	-11.4365
+▁innocent	-11.4365
+▁creation	-11.4365
+▁Week	-11.4366
+▁Royal	-11.4367
+giving	-11.4414
+▁corrupt	-11.4495
+▁atmosphere	-11.4526
+▁easiest	-11.4526
+▁Recipe	-11.4526
+▁minimum	-11.4526
+▁clown	-11.4526
+▁sneak	-11.4526
+▁covid	-11.4527
+▁smash	-11.4527
+▁Listen	-11.4527
+▁Bridge	-11.4528
+▁reserv	-11.4532
+▁accountant	-11.4539
+▁coope	-11.4564
+responsi	-11.469
+▁International	-11.469
+▁Service	-11.469
+▁disgusting	-11.469
+▁rabbit	-11.469
+▁senior	-11.469
+▁kingdom	-11.469
+▁almond	-11.469
+▁Quaran	-11.469
+▁combine	-11.469
+▁Center	-11.4691
+▁convey	-11.4691
+awareness	-11.4691
+spoken	-11.4769
+ibly	-11.4785
+▁satisfy	-11.4857
+responsibilities	-11.4857
+▁agriculture	-11.4857
+▁colleague	-11.4857
+▁marijuana	-11.4857
+▁pollution	-11.4857
+▁shampoo	-11.4857
+▁consistent	-11.4857
+▁supply	-11.4857
+▁depressed	-11.4857
+▁retrieve	-11.4857
+▁manipulat	-11.4857
+▁fortune	-11.4857
+▁jeans	-11.4858
+▁afterwards	-11.4858
+▁initial	-11.4858
+▁float	-11.4859
+functional	-11.4921
+ATE	-11.495
+▁elephant	-11.5026
+▁luxury	-11.5026
+▁physics	-11.5026
+▁fulfill	-11.5026
+solving	-11.5026
+▁ISIS	-11.5026
+▁NBA	-11.5026
+▁slept	-11.5026
+nigger	-11.5027
+▁lottery	-11.5028
+▁giant	-11.5028
+▁Halloween	-11.5199
+▁Thursday	-11.5199
+▁analyzing	-11.5199
+▁awkward	-11.5199
+▁dinosaur	-11.5199
+▁emergency	-11.5199
+▁enemies	-11.5199
+▁starving	-11.5199
+▁tablespoon	-11.5199
+NSFW	-11.5199
+▁WHO	-11.5199
+▁database	-11.5199
+▁landmarks	-11.5199
+▁predator	-11.5199
+▁bullet	-11.5199
+▁lyrics	-11.5199
+▁sector	-11.5199
+▁costume	-11.5199
+▁reject	-11.52
+▁upgrade	-11.5203
+▁constant	-11.5361
+▁construct	-11.5369
+▁Catholic	-11.5374
+▁administration	-11.5374
+▁astronaut	-11.5374
+▁circle	-11.5374
+▁desperate	-11.5374
+▁prioritize	-11.5374
+▁teaspoon	-11.5374
+▁cocaine	-11.5374
+▁interpret	-11.5374
+▁Greece	-11.5374
+▁insightful	-11.5374
+▁complement	-11.5374
+▁preparing	-11.5374
+▁casual	-11.5374
+▁drain	-11.5374
+▁advise	-11.5375
+▁purple	-11.5375
+▁colour	-11.5375
+▁divide	-11.5375
+Peace	-11.5375
+govern	-11.5386
+▁slide	-11.539
+▁Latin	-11.5405
+▁dimension	-11.5514
+▁McDonald	-11.5553
+▁contribution	-11.5553
+▁hockey	-11.5553
+▁pencil	-11.5553
+▁scenario	-11.5553
+▁ultimate	-11.5553
+▁cloves	-11.5553
+▁fridge	-11.5553
+▁Second	-11.5553
+▁Senator	-11.5553
+▁rescue	-11.5553
+▁Daniel	-11.5554
+▁sample	-11.5554
+Company	-11.5554
+▁plug	-11.5554
+▁compete	-11.5561
+volving	-11.5735
+▁Children	-11.5735
+▁victory	-11.5735
+▁villain	-11.5735
+▁Museum	-11.5735
+▁urban	-11.5735
+▁Western	-11.5735
+▁Emma	-11.5738
+▁skate	-11.5742
+reaching	-11.5748
+▁severe	-11.5753
+▁featur	-11.5783
+[	-11.5817
+▁Otherwise	-11.592
+▁assignment	-11.592
+▁coconut	-11.592
+▁manufacture	-11.592
+▁patience	-11.592
+▁puppy	-11.592
+▁puzzle	-11.592
+▁stereotype	-11.592
+▁tactics	-11.592
+‍	-11.592
+▁narrow	-11.592
+▁Knock	-11.592
+▁drank	-11.592
+▁Space	-11.592
+▁actress	-11.592
+▁estimate	-11.592
+▁Boston	-11.592
+▁Wood	-11.592
+▁gaming	-11.592
+▁steam	-11.5921
+▁wander	-11.5921
+▁depend	-11.5956
+maybe	-11.6035
+▁integration	-11.6108
+▁Seattle	-11.6108
+▁Windows	-11.6108
+▁volume	-11.6108
+▁worship	-11.6108
+▁cruel	-11.6108
+▁sentiment	-11.6108
+▁concise	-11.6108
+@	-11.7322
+<	-11.7448
+£	-11.846
+°	-11.9473
+ا	-12.0307
+🙄	-12.0904
+🏼	-12.1216
+•	-12.1539
+😊	-12.2574
+á	-12.3329
+ó	-12.3329
+🍗	-12.3329
+💕	-12.3329
+🏻	-12.3729
+ل	-12.4146
+♀	-12.4581
+✨	-12.4581
+í	-12.5035
+ر	-12.5035
+🤔	-12.5512
+🤷	-12.5512
+😍	-12.6538
+♂	-12.7093
+🥴	-12.7093
+®	-12.7682
+👌	-12.7682
+😁	-12.7682
+“	-12.8183
+}	-12.8307
+ی	-12.8307
+💖	-12.8307
+💙	-12.8307
+🙌	-12.8307
+م	-12.8973
+🤦	-12.8973
+	-12.9688
+👏	-12.9688
+😀	-12.9688
+è	-12.9688
+😔	-13.0457
+😢	-13.0457
+🤗	-13.0457
+ب	-13.0457
+ō	-13.129
+س	-13.129
+ہ	-13.129
+😎	-13.129
+😬	-13.129
+👇	-13.2199
+🕊	-13.2199
+😩	-13.2199
+ک	-13.3199
+♥	-13.3199
+🔥	-13.3199
+à	-13.431
+د	-13.431
+و	-13.431
+ے	-13.431
+✌	-13.431
+🏽	-13.431
+💫	-13.431
+😅	-13.431
+😇	-13.431
+🥰	-13.431
+ö	-13.431
+ن	-13.431
+ي	-13.431
+ع	-13.556
+💀	-13.556
+ã	-13.556
+ä	-13.556
+̛	-13.556
+€	-13.556
+☕	-13.556
+👑	-13.556
+😏	-13.556
+😜	-13.556
+ı	-13.6989
+♡	-13.6989
+➡	-13.6989
+🐝	-13.6989
+👊	-13.6989
+💓	-13.6989
+💪	-13.6989
+😋	-13.6989
+😷	-13.6989
+ت	-13.6989
+â	-13.6989
+̨	-13.8656
+‬	-13.8656
+⭐	-13.8656
+🌺	-13.8656
+👆	-13.8656
+😒	-13.8656
+😤	-13.8656
+🛰	-13.8656
+🥂	-13.8656
+😫	-14.0656
+×	-14.0656
+α	-14.0656
+ق	-14.0656
+⁩	-14.0656
+✍	-14.0656
+➦	-14.0656
+。	-14.0656
+🌷	-14.0656
+😃	-14.0656
+😝	-14.0656
+😪	-14.0656
+🙋	-14.0656
+🤓	-14.0656
+🧿	-14.0656
+ح	-14.0656
+ल	-14.3155
+¢	-14.3156
+ì	-14.3156
+⁣	-14.3156
+☝	-14.3156
+☠	-14.3156
+⚡	-14.3156
+✔	-14.3156
+🇦	-14.3156
+🌐	-14.3156
+🏖	-14.3156
+👋	-14.3156
+💐	-14.3156
+📖	-14.3156
+😌	-14.3156
+ج	-14.3156
+🙃	-14.3156
+ि	-14.6488
+Â	-14.6489
+ذ	-14.6489
+»	-14.6489
+ā	-14.6489
+μ	-14.6489
+ة	-14.6489
+✈	-14.6489
+✡	-14.6489
+❄	-14.6489
+❣	-14.6489
+ー	-14.6489
+🇮	-14.6489
+🌉	-14.6489
+🌬	-14.6489
+🎂	-14.6489
+🏵	-14.6489
+👅	-14.6489
+💋	-14.6489
+💝	-14.6489
+💟	-14.6489
+💥	-14.6489
+🔔	-14.6489
+😚	-14.6489
+😺	-14.6489
+🚩	-14.6489
+🚶	-14.6489
+🤩	-14.6489
+🤬	-14.6489
+🤮	-14.6489
+🥁	-14.6489
+🧡	-14.6489
+û	-14.6489
+る	-14.6489
+を	-14.6489
+の	-14.6489
+्	-15.1487
+ा	-15.1489
+ğ	-15.1489
+ط	-15.1489
+田	-15.1489
+🥚	-15.1489
+Ç	-15.1489
+Ü	-15.1489
+æ	-15.1489
+ð	-15.1489
+ú	-15.1489
+ă	-15.1489
+š	-15.1489
+ż	-15.1489
+ə	-15.1489
+؟	-15.1489
+آ	-15.1489
+ز	-15.1489
+ك	-15.1489
+ى	-15.1489
+پ	-15.1489
+ں	-15.1489
+ھ	-15.1489
+۔	-15.1489
+ी	-15.1489
+ु	-15.1489
+ง	-15.1489
+□	-15.1489
+◍	-15.1489
+☃	-15.1489
+♾	-15.1489
+⛪	-15.1489
+✩	-15.1489
+❓	-15.1489
+、	-15.1489
+🇧	-15.1489
+🇭	-15.1489
+🇱	-15.1489
+🇴	-15.1489
+🌄	-15.1489
+🌅	-15.1489
+🌊	-15.1489
+🌌	-15.1489
+🌞	-15.1489
+🌠	-15.1489
+🌤	-15.1489
+🌱	-15.1489
+🍸	-15.1489
+🍾	-15.1489
+🎀	-15.1489
+🎈	-15.1489
+🎙	-15.1489
+🎸	-15.1489
+🐒	-15.1489
+🐦	-15.1489
+🐷	-15.1489
+👈	-15.1489
+👣	-15.1489
+👻	-15.1489
+💃	-15.1489
+💅	-15.1489
+💌	-15.1489
+💎	-15.1489
+💒	-15.1489
+💣	-15.1489
+💦	-15.1489
+💩	-15.1489
+💰	-15.1489
+📢	-15.1489
+📸	-15.1489
+🗡	-15.1489
+😖	-15.1489
+😹	-15.1489
+🙁	-15.1489
+🙊	-15.1489
+🚗	-15.1489
+🤍	-15.1489
+🤑	-15.1489
+🤞	-15.1489
+🥇	-15.1489
+🧁	-15.1489
+👁	-15.1489
+å	-15.1489
+ο	-15.1489
+خ	-15.1489
+َ	-15.1489
+ɪ	-15.1489
+न	-15.1489
+อ	-15.1489
+ं	-15.149
+🦋	-16.1409
+🥑	-16.141
+🤨	-16.1411
+🕺	-16.1412
+🐾	-16.1413
+🎥	-16.1414
+🍹	-16.1415
+愛	-16.1416
+☂	-16.1417
+√	-16.1418
+⃣	-16.1419
+स	-16.142
+ş	-16.1421
+É	-16.1422
+¡	-16.1423
+🧐	-16.1424
+🤢	-16.1425
+🤝	-16.1426
+😣	-16.1427
+😛	-16.1428
+😈	-16.1429
+🗣	-16.143
+💨	-16.1431
+今	-16.1432
+⚠	-16.1433
+ف	-16.1434
+œ	-16.1435
+🤘	-16.1436
+😠	-16.1437
+😞	-16.1438
+🌿	-16.1439
+✅	-16.144
+⚽	-16.1441
+☀	-16.1442
+π	-16.1443
+ï	-16.1444
+😄	-16.1445
+💔	-16.1446
+🇨	-16.1447
+⠀	-16.1448
+‪	-16.1449
+😻	-16.145
+😡	-16.1451
+😑	-16.1452
+🎉	-16.1453
+⁦	-16.1454
+😳	-16.1455
+☺	-16.1456
+😱	-16.1457
+👉	-16.1458
+🎶	-16.1459
+🌹	-16.146
+🌟	-16.1461
+🙂	-16.1462
+💛	-16.1463
+👀	-16.1464
+🍀	-16.1465
+🤡	-16.1466
+̊	-16.1467
+🙈	-16.1468
+¿	-16.1469
+🥺	-16.147
+💚	-16.1471
+💗	-16.1472
+ๆ	-16.1473
+⁄	-16.1474
+😘	-16.1475
+💯	-16.1476
+😆	-16.1477
+🏾	-16.1478
+ü	-16.1479
+ç	-16.148
+🇸	-16.1481
+́	-16.1482
+💜	-16.1483
+😉	-16.1484
+🇺	-16.1485
+ठ	-16.1486
+👍	-16.1486
+ट	-16.1487
+💞	-16.1487
+ñ	-16.1488
+ओ	-16.1488
+ज	-16.1488
+प	-16.1488
+ब	-16.1488
+ω	-16.1489
+ش	-16.1489
+−	-16.1489
+­	-16.1489
+ê	-16.1489
+≠	-16.1489
+❶	-16.1489
+💆	-16.1489
+€	-16.1489
+©	-16.1489
+«	-16.1489
+¶	-16.1489
+À	-16.1489
+Í	-16.1489
+ß	-16.1489
+î	-16.1489
+÷	-16.1489
+ù	-16.1489
+ÿ	-16.1489
+ĝ	-16.1489
+ź	-16.1489
+ɒ	-16.1489
+ʻ	-16.1489
+Ξ	-16.1489
+Π	-16.1489
+ν	-16.1489
+υ	-16.1489
+е	-16.1489
+כ	-16.1489
+ן	-16.1489
+ض	-16.1489
+ظ	-16.1489
+ه	-16.1489
+ّ	-16.1489
+ٹ	-16.1489
+گ	-16.1489
+ۓ	-16.1489
+द	-16.1489
+น	-16.1489
+บ	-16.1489
+ั	-16.1489
+้	-16.1489
+ᴗ	-16.1489
+₤	-16.1489
+₹	-16.1489
+←	-16.1489
+⌚	-16.1489
+⌨	-16.1489
+⏰	-16.1489
+▶	-16.1489
+●	-16.1489
+☑	-16.1489
+☻	-16.1489
+♨	-16.1489
+♪	-16.1489
+♫	-16.1489
+⚘	-16.1489
+⚰	-16.1489
+⚾	-16.1489
+✝	-16.1489
+✳	-16.1489
+❇	-16.1489
+❗	-16.1489
+】	-16.1489
+あ	-16.1489
+こ	-16.1489
+さ	-16.1489
+す	-16.1489
+で	-16.1489
+と	-16.1489
+了	-16.1489
+分	-16.1489
+士	-16.1489
+夢	-16.1489
+寵	-16.1489
+強	-16.1489
+心	-16.1489
+扎	-16.1489
+況	-16.1489
+洗	-16.1489
+禅	-16.1489
+紳	-16.1489
+練	-16.1489
+門	-16.1489
+나	-16.1489
+러	-16.1489
+브	-16.1489
+인	-16.1489
+	-16.1489
+︎	-16.1489
+🃏	-16.1489
+🆒	-16.1489
+🆓	-16.1489
+🇩	-16.1489
+🇪	-16.1489
+🇬	-16.1489
+🇳	-16.1489
+🇹	-16.1489
+🇼	-16.1489
+🇿	-16.1489
+🌀	-16.1489
+🌃	-16.1489
+🌆	-16.1489
+🌈	-16.1489
+🌍	-16.1489
+🌕	-16.1489
+🌙	-16.1489
+🌸	-16.1489
+🌻	-16.1489
+🍁	-16.1489
+🍄	-16.1489
+🍊	-16.1489
+🍑	-16.1489
+🍟	-16.1489
+🍦	-16.1489
+🍭	-16.1489
+🍷	-16.1489
+🍻	-16.1489
+🍽	-16.1489
+🎁	-16.1489
+🎄	-16.1489
+🎆	-16.1489
+🎇	-16.1489
+🎓	-16.1489
+🎗	-16.1489
+🎢	-16.1489
+🎤	-16.1489
+🎨	-16.1489
+🎪	-16.1489
+🎲	-16.1489
+🎵	-16.1489
+🎼	-16.1489
+🏆	-16.1489
+🏏	-16.1489
+🏔	-16.1489
+🏘	-16.1489
+🏚	-16.1489
+🏠	-16.1489
+🏤	-16.1489
+🏴	-16.1489
+🏿	-16.1489
+🐀	-16.1489
+🐂	-16.1489
+🐃	-16.1489
+🐄	-16.1489
+🐌	-16.1489
+🐘	-16.1489
+🐚	-16.1489
+🐠	-16.1489
+🐧	-16.1489
+🐭	-16.1489
+🐮	-16.1489
+🐯	-16.1489
+🐱	-16.1489
+👄	-16.1489
+👎	-16.1489
+👐	-16.1489
+👗	-16.1489
+👨	-16.1489
+👮	-16.1489
+👶	-16.1489
+👿	-16.1489
+💉	-16.1489
+💘	-16.1489
+💡	-16.1489
+💭	-16.1489
+💼	-16.1489
+📷	-16.1489
+📺	-16.1489
+🔊	-16.1489
+🔓	-16.1489
+🔗	-16.1489
+🔫	-16.1489
+🔹	-16.1489
+🕉	-16.1489
+🕯	-16.1489
+🕵	-16.1489
+🖋	-16.1489
+🖖	-16.1489
+🖥	-16.1489
+🗑	-16.1489
+😐	-16.1489
+😕	-16.1489
+😧	-16.1489
+😰	-16.1489
+😴	-16.1489
+😵	-16.1489
+🚀	-16.1489
+🚌	-16.1489
+🚘	-16.1489
+🚴	-16.1489
+🛣	-16.1489
+🤖	-16.1489
+🤜	-16.1489
+🤟	-16.1489
+🤠	-16.1489
+🤤	-16.1489
+🤪	-16.1489
+🤭	-16.1489
+🤲	-16.1489
+🤸	-16.1489
+🥒	-16.1489
+🥬	-16.1489
+🥭	-16.1489
+🥲	-16.1489
+🥵	-16.1489
+🦁	-16.1489
+🦃	-16.1489
+🦄	-16.1489
+🦌	-16.1489
+🦜	-16.1489
+🧀	-16.1489
+🧬	-16.1489
+🧳	-16.1489
+🙏	-16.1489
+ก	-16.1489
+λ	-16.1489
+أ	-16.1489
+ص	-16.1489
+ῖ	-16.1489
+【	-16.1489
+状	-16.1489
+自	-16.1489
+🧖	-16.1489
+ˈ	-16.1489
+ป	-16.1489
+ร	-16.1489
+ห	-16.1489
+ู	-16.1489
+โ	-16.1489
+い	-16.1489
+っ	-16.1489
+て	-16.1489
+も	-16.1489
+持	-16.1489
+ม	-16.1489
+ะ	-16.1489
+ึ	-16.1489
+क	-16.149
+ग	-16.149
+ड	-16.149
+ह	-16.149
+े	-16.149
+่	-16.149
+🤣	-16.149

chat_sprint_artifacts/tinygpt.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37118ed4c7497877ccaf6fafebeb63ed37d4988e8f414cac27ae55d31dd43b02
+size 59904321

chat_sprint_artifacts/train_ids.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f388f53957e50f65bd632083829f7f9cb198ef32867d4bb8bdaf0377fd7e698e
+size 48539814

chat_sprint_artifacts/val_ids.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94eec301757b67b5861153d442c8e8ac5f9d67753d264928c1a7cc374deea455
+size 1502364

chat_sprint_artifacts/vocab_size.txt

@@ -0,0 +1 @@
+5000

train.py

@@ -0,0 +1,773 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+chat_sprint_standalone.py
+One-file pipeline: collect datasets -> reformat as You:/Bot: -> train tiny GPT (CUDA) -> sample & save model
+
+Requirements (install once):
+  pip install torch datasets sentencepiece tqdm numpy
+
+Run:
+  python chat_sprint_standalone.py
+"""
+
+import os, re, time, math, random, json
+from pathlib import Path
+from typing import List, Optional, Tuple
+from itertools import islice
+from contextlib import nullcontext
+
+import numpy as np
+from tqdm import tqdm
+from datasets import load_dataset, get_dataset_config_names
+import sentencepiece as spm
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# --------------------------
+# Global config (tweak here)
+# --------------------------
+SEED = 1337
+random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+if DEVICE.type == "cuda":
+    torch.set_float32_matmul_precision("high")
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+SAVE_DIR = Path("./chat_sprint_artifacts")
+SAVE_DIR.mkdir(parents=True, exist_ok=True)
+
+# How many formatted You:/Bot: pairs to KEEP from each dataset (first run)
+CAPS = {
+    "shakespeare": 15000,
+    "jokes":       20000,
+    "dadjokes":     8000,
+    "rsarcasm":     8000,   # Thewillonline/reddit-sarcasm
+    "figlang":     12000,
+    "shower":       4000,   # HuggingFaceGECLM/REDDIT_submissions split "Showerthoughts"
+    "personas":     2000,
+    "tweeteval":    4000,
+    "fourchan":      500,   # dataset is tiny (~195)
+    "elonvtrump":   3000,
+}
+
+# Upper bound on rows to SCAN in each streaming dataset
+SCAN = {
+    "jokes":       120_000,
+    "dadjokes":     60_000,
+    "rsarcasm":    120_000,
+    "figlang":     150_000,
+    "shower":      250_000,
+    "personas":     30_000,
+    "tweeteval":    60_000,
+    "fourchan":      2_000,
+    "elonvtrump":   60_000,
+}
+
+MAX_TOTAL_PAIRS = 60_000
+MAX_LEN = 120
+
+# Tokenizer
+VOCAB_SIZE = 1500
+TOKENIZER_PREFIX = SAVE_DIR / "spm_chat"
+USER_SYMBOLS = ["You:", "Bot:", "[STYLE=Snark]", "[FORM=TWEET]", "[FORM=HEADLINE]", "[MOOD=Unhinged]", "[MOOD=Cheeky]"]
+
+# Model size & train budget (~5 minutes on RTX 3090 with default)
+block_size   = 256
+n_layer      = 6
+n_head       = 6
+n_embd       = 384   # ~11.9M params
+dropout      = 0.0
+
+MAX_SECONDS  = 300   # hard cap
+train_steps  = 5000  # big number; time cap will stop near 5 min
+log_interval = 100
+eval_every   = 400
+batch_size   = 24
+accum_steps  = 3
+base_lr      = 3e-3
+min_lr       = 5e-4
+warmup_ratio = 0.06
+
+# Sampling defaults
+TEMP = 0.8
+TOP_K = 60
+TOP_P = 0.95
+REP_PEN = 1.08
+
+# --------------------------
+# Helpers: cleaning & format
+# --------------------------
+URL_RE   = re.compile(r"https?://\S+|www\.\S+", re.IGNORECASE)
+MENT_RE  = re.compile(r"@\w+")
+WS_RE    = re.compile(r"\s+")
+QUOTE_RE = re.compile(r"^[\"'“”‘’]+|[\"'“”‘’]+$")
+
+def clean_text(s: str) -> str:
+    s = s.strip()
+    s = URL_RE.sub("", s)
+    s = MENT_RE.sub("", s)
+    s = QUOTE_RE.sub("", s)
+    s = WS_RE.sub(" ", s)
+    return s.strip()
+
+def shorten_to(s: str, n: int) -> str:
+    s = re.sub(r"\s+", " ", s.strip())
+    if len(s) <= n: return s
+    cut = max(s.rfind(". ", 0, n), s.rfind("! ", 0, n), s.rfind("? ", 0, n))
+    if cut != -1: return s[:cut+1].strip()
+    return s[:n].strip()
+
+def keep_or_clip(s: str, min_len: int = 6, max_len: int = MAX_LEN) -> Optional[str]:
+    if not s: return None
+    s = re.sub(r"\s+", " ", s.strip())
+    if len(s) < min_len: return None
+    return shorten_to(s, max_len)
+
+def turn(you: str, bot: str, tags: str = "") -> str:
+    lines = [f"You: {you}".rstrip()]
+    if tags: lines.append(tags)
+    lines.append(f"Bot: {bot}".rstrip())
+    lines.append("")
+    return "\n".join(lines)
+
+def limited(ds, limit: int):
+    try:
+        return ds.take(limit)
+    except Exception:
+        return islice(ds, limit)
+
+def get_first_nonempty(ex, keys) -> Optional[str]:
+    for k in keys:
+        v = ex.get(k)
+        if isinstance(v, str) and v.strip():
+            return v
+    return None
+
+# --------------------------
+# Collectors (console tqdm)
+# --------------------------
+def collect_shakespeare(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset(
+            "text",
+            data_files={"train": "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"},
+            split="train",
+            streaming=True,
+        )
+        kept = 0
+        pbar = tqdm(total=CAPS["shakespeare"], desc="[shakespeare]", unit="pair", leave=False, ncols=100)
+        for row in ds:
+            txt = keep_or_clip(clean_text(row["text"]))
+            if not txt: continue
+            pairs.append(turn("Continue in Shakespearean style.", txt))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= CAPS["shakespeare"]: break
+        pbar.close()
+        print(f"[ok] shakespeare kept={kept}")
+    except Exception as e:
+        print(f"[skip] shakespeare: {e}")
+
+def collect_reddit_jokes(pairs: List[str], overall: tqdm):
+    for dsid in ["SocialGrep/one-million-reddit-jokes", "SocialGrep/reddit_jokes", "timc1/reddit_jokes"]:
+        try:
+            ds = load_dataset(dsid, split="train", streaming=True)
+            kept = 0
+            pbar = tqdm(total=CAPS["jokes"], desc="[jokes]", unit="pair", leave=False, ncols=100)
+            for ex in limited(ds, SCAN["jokes"]):
+                title = keep_or_clip(clean_text(str(ex.get("title") or "")))
+                body  = keep_or_clip(clean_text(str(ex.get("selftext") or ex.get("body") or "")))
+                if body and title:
+                    pairs.append(turn(title, body))
+                elif title:
+                    pairs.append(turn("Tell me a short joke.", title))
+                else:
+                    continue
+                kept += 1; pbar.update(1); overall.update(1)
+                if kept >= CAPS["jokes"]: break
+            pbar.close()
+            print(f"[ok] jokes {dsid} kept={kept}")
+            return
+        except Exception as e:
+            print(f"[try next] jokes {dsid}: {e}")
+    print("[skip] jokes: none worked")
+
+def collect_dadjokes(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("shuttie/reddit-dadjokes", split="train", streaming=True)
+        kept = 0
+        pbar = tqdm(total=CAPS["dadjokes"], desc="[dadjokes]", unit="pair", leave=False, ncols=100)
+        for ex in limited(ds, SCAN["dadjokes"]):
+            setup = keep_or_clip(clean_text(str(ex.get("setup") or ex.get("instruction") or ex.get("input") or "")))
+            punch = keep_or_clip(clean_text(str(ex.get("punchline") or ex.get("output") or "")))
+            if not (setup and punch): continue
+            pairs.append(turn(setup, punch))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= CAPS["dadjokes"]: break
+        pbar.close()
+        print(f"[ok] dadjokes kept={kept}")
+    except Exception as e:
+        print(f"[skip] dadjokes: {e}")
+
+def collect_reddit_sarcasm(pairs: List[str], overall: tqdm):
+    """Thewillonline/reddit-sarcasm — flexible parsing, scan+keep bars."""
+    try:
+        ds = load_dataset("Thewillonline/reddit-sarcasm", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["rsarcasm"], SCAN["rsarcasm"]
+        scanbar = tqdm(total=scan_cap, desc="[sarcasm scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[sarcasm kept]", unit="pair", leave=False, ncols=100)
+
+        PATS = [
+            re.compile(r"User\s*:\s*(.+?)\s*(?:Reddit\s*Comment|Comment|Reply)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
+            re.compile(r"Post\s*:\s*(.+?)\s*(?:Top\s*Comment|Comment)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
+        ]
+        def parse(raw: str) -> Tuple[Optional[str], Optional[str]]:
+            raw = raw.replace("<|endoftext|>", "\n")
+            for pat in PATS:
+                m = pat.search(raw)
+                if m:
+                    return m.group(1).strip(), m.group(2).strip()
+            lines = [ln.strip() for ln in raw.splitlines() if ln.strip()]
+            if len(lines) >= 2: return lines[0], lines[1]
+            if len(lines) == 1: return "Reply with sarcasm:", lines[0]
+            return None, None
+
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            you, bot = parse(str(ex.get("text") or ""))
+            you = keep_or_clip(you); bot = keep_or_clip(bot)
+            if you and bot:
+                pairs.append(turn(you, bot, "[STYLE=Snark]"))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+            if scanned % 2000 == 0:
+                keepbar.set_postfix(rate=f"{kept/max(1,scanned):.2%}")
+        scanbar.close(); keepbar.close()
+        print(f"[ok] reddit-sarcasm kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] reddit-sarcasm: {e}")
+
+def collect_figlang(pairs: List[str], overall: tqdm):
+    for dsid in ["tasksource/figlang2020-sarcasm", "tasksource/figlang2020_sarcasm"]:
+        try:
+            ds = load_dataset(dsid, split="train", streaming=True)
+            kept = 0
+            pbar = tqdm(total=CAPS["figlang"], desc="[figlang]", unit="pair", leave=False, ncols=100)
+            for ex in limited(ds, SCAN["figlang"]):
+                ctx = ex.get("context")
+                if isinstance(ctx, list) and ctx:
+                    context_str = " ".join(str(c) for c in ctx[-2:])
+                else:
+                    context_str = str(ex.get("context") or ex.get("prompt") or "")
+                reply = str(ex.get("response") or ex.get("answer") or ex.get("text") or "")
+                context_str = keep_or_clip(clean_text(context_str))
+                reply      = keep_or_clip(clean_text(reply))
+                if reply:
+                    if context_str:
+                        pairs.append(turn(context_str, reply, "[STYLE=Snark]"))
+                    else:
+                        pairs.append(turn("Reply with sarcasm:", reply, "[STYLE=Snark]"))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= CAPS["figlang"]: break
+            pbar.close()
+            print(f"[ok] figlang {dsid} kept={kept}")
+            return
+        except Exception as e:
+            print(f"[try next] figlang {dsid}: {e}")
+    print("[skip] figlang")
+
+def collect_showerthoughts(pairs: List[str], overall: tqdm):
+    """Use REEDIT_submissions split 'Showerthoughts' directly (no 'train')."""
+    try:
+        ds = load_dataset("HuggingFaceGECLM/REDDIT_submissions", split="Showerthoughts", streaming=True)
+        keep_cap, scan_cap = CAPS["shower"], SCAN["shower"]
+        scanbar = tqdm(total=scan_cap, desc="[shower scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[shower kept]", unit="pair", leave=False, ncols=100)
+
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            title = get_first_nonempty(ex, ["title", "selftext", "text"]) or ""
+            text = keep_or_clip(clean_text(title))
+            if text:
+                pairs.append(turn("Give me a shower thought.", text))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+        scanbar.close(); keepbar.close()
+        print(f"[ok] showerthoughts kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] showerthoughts: {e}")
+
+def collect_personas(pairs: List[str], overall: tqdm):
+    """Non-streaming is more reliable for this dataset."""
+    try:
+        ds = load_dataset("NapthaAI/twitter_personas")["train"]
+        keep_cap = CAPS["personas"]
+        pbar = tqdm(total=keep_cap, desc="[personas]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in ds:
+            desc = get_first_nonempty(ex, ["description","persona","bio","text","content","full_text"])
+            if not isinstance(desc, str) and isinstance(ex.get("content"), dict):
+                desc = ex["content"].get("text")
+            desc = keep_or_clip(clean_text(str(desc or "")))
+            if not desc: continue
+            pairs.append(turn("Adopt this persona in one sentence.", desc, "[FORM=TWEET]"))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] personas kept={kept}")
+    except Exception as e:
+        print(f"[skip] personas: {e}")
+
+def collect_tweeteval(pairs: List[str], overall: tqdm):
+    """Handle super_tweeteval (text_1/text_2, etc.) and fallback tweet_eval."""
+    def extract_pair(ex):
+        t = ex.get("text")
+        if isinstance(t, str) and t.strip():
+            return "React with a sharp one-liner.", t
+        for a,b in [("text_1","text_2"), ("sentence1","sentence2"),
+                    ("premise","hypothesis"), ("question","answer"),
+                    ("context","response"), ("tweet1","tweet2")]:
+            t1, t2 = ex.get(a), ex.get(b)
+            if isinstance(t1, str) and t1.strip() and isinstance(t2, str) and t2.strip():
+                return t1, t2
+        return None
+
+    def run_on(dsname, pick, is_super):
+        keep_cap, scan_cap = CAPS["tweeteval"], SCAN["tweeteval"]
+        pbar = tqdm(total=keep_cap, desc=f"[tweeteval:{pick}]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        ds = load_dataset(dsname, pick, split="train", streaming=True)
+        for ex in limited(ds, scan_cap):
+            pair = extract_pair(ex) if is_super else ("React with a sharp one-liner.", ex.get("text")) if ex.get("text") else None
+            if not pair: continue
+            you, bot = pair
+            you = keep_or_clip(clean_text(str(you or ""))); bot = keep_or_clip(clean_text(str(bot or "")))
+            if not (you and bot): continue
+            tag = "[STYLE=Snark]" if you and you != "React with a sharp one-liner." else ""
+            pairs.append(turn(you, bot, tag))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        return kept
+
+    kept_total = 0
+    try:
+        cfgs = get_dataset_config_names("cardiffnlp/super_tweeteval")
+    except Exception:
+        cfgs = []
+    prio = ["irony","sarcasm","humor","sentiment","emoji","emotion","stance","offensive","hate"]
+    ordered = [c for c in prio if c in cfgs] + [c for c in cfgs if c not in prio]
+    for pick in ordered:
+        kept_total += run_on("cardiffnlp/super_tweeteval", pick, True)
+        if kept_total >= CAPS["tweeteval"]:
+            print(f"[ok] tweeteval(super) kept={kept_total}"); return
+    if kept_total == 0:
+        try:
+            base_cfgs = get_dataset_config_names("cardiffnlp/tweet_eval")
+        except Exception:
+            base_cfgs = []
+        ordered_b = [c for c in prio if c in base_cfgs] + [c for c in base_cfgs if c not in prio]
+        for pick in ordered_b:
+            kept_total += run_on("cardiffnlp/tweet_eval", pick, False)
+            if kept_total >= CAPS["tweeteval"]:
+                print(f"[ok] tweeteval(base) kept={kept_total}"); return
+    print(f"[ok] tweeteval kept={kept_total}")
+
+def collect_fourchan(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("sbussiso/4chan-dataset", split="train", streaming=True)
+        keep_cap = min(CAPS["fourchan"], 195)
+        pbar = tqdm(total=keep_cap, desc="[4chan]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in limited(ds, SCAN["fourchan"]):
+            prompt = keep_or_clip(clean_text(str(ex.get("prompt") or "")))
+            resp   = keep_or_clip(clean_text(str(ex.get("response") or "")))
+            if prompt and resp:
+                pairs.append(turn(prompt, resp))
+                kept += 1; pbar.update(1); overall.update(1)
+            else:
+                txt = keep_or_clip(clean_text(str(ex.get("text") or ex.get("body") or ex.get("content") or "")))
+                if txt:
+                    pairs.append(turn("Drop a spicy one-liner.", txt))
+                    kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] 4chan kept={kept}")
+    except Exception as e:
+        print(f"[skip] 4chan: {e}")
+
+def collect_elon_trump(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("MasaFoundation/Twitter_X_Elon_vs_Trump", split="train", streaming=True, revision="refs/convert/parquet")
+        keep_cap, scan_cap = CAPS["elonvtrump"], SCAN["elonvtrump"]
+        scanbar = tqdm(total=scan_cap, desc="[elon_vs_trump scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[elon_vs_trump kept]", unit="pair", leave=False, ncols=100)
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            txt = get_first_nonempty(ex, ["content","text","tweet","full_text"]) or ""
+            txt = keep_or_clip(clean_text(txt))
+            if txt:
+                pairs.append(turn("[FORM=TWEET] One sentence hot take:", txt, "[FORM=TWEET]"))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+        scanbar.close(); keepbar.close()
+        print(f"[ok] Elon_vs_Trump kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] Elon_vs_Trump: {e}")
+
+# --------------------------
+# Build corpus
+# --------------------------
+def build_corpus() -> Path:
+    pairs: List[str] = []
+    total_target = sum(CAPS.values())
+    print("[1/6] Collecting & reformatting datasets (streaming, capped)…")
+    overall = tqdm(total=total_target, desc="[all] collecting", unit="pair", ncols=100)
+
+    collectors = [
+        collect_shakespeare,
+        collect_reddit_jokes,
+        collect_dadjokes,
+        collect_reddit_sarcasm,
+        collect_figlang,
+        collect_showerthoughts,
+        collect_personas,
+        collect_tweeteval,
+        collect_fourchan,
+        collect_elon_trump,
+    ]
+    for fn in collectors:
+        try:
+            fn(pairs, overall)
+        except Exception as e:
+            print(f"[collector error] {fn.__name__}: {e}")
+
+    overall.close()
+    print("[2/6] Deduplicating & clipping…")
+    seen = set(); deduped = []
+    for block in pairs:
+        try:
+            bot_line = [ln for ln in block.splitlines() if ln.startswith("Bot:")][0]
+            key = bot_line[4:].strip().lower()
+        except Exception:
+            key = block.strip().lower()
+        if key in seen: continue
+        seen.add(key); deduped.append(block)
+
+    random.shuffle(deduped)
+    if len(deduped) > MAX_TOTAL_PAIRS:
+        deduped = deduped[:MAX_TOTAL_PAIRS]
+
+    out_path = SAVE_DIR / "corpus.txt"
+    out_path.write_text("\n".join(deduped), encoding="utf-8")
+    print(f"     wrote {len(deduped)} pairs → {out_path}")
+    return out_path
+
+# --------------------------
+# SentencePiece tokenizer
+# --------------------------
+def train_spm(corpus_path: Path) -> spm.SentencePieceProcessor:
+    print("[3/6] Training SentencePiece tokenizer…")
+    spm.SentencePieceTrainer.Train(
+        input=str(corpus_path),
+        model_prefix=str(TOKENIZER_PREFIX),
+        vocab_size=VOCAB_SIZE,
+        model_type="unigram",
+        character_coverage=1.0,
+        user_defined_symbols=USER_SYMBOLS,
+        bos_id=1, eos_id=2, unk_id=0, pad_id=-1
+    )
+    sp = spm.SentencePieceProcessor()
+    sp.load(f"{TOKENIZER_PREFIX}.model")
+    print(f"     tokenizer saved at {TOKENIZER_PREFIX}.model")
+    return sp
+
+# --------------------------
+# Encode to token IDs
+# --------------------------
+def encode_corpus_to_ids(sp: spm.SentencePieceProcessor, corpus_path: Path):
+    print("[4/6] Encoding corpus to token IDs…")
+    text = corpus_path.read_text(encoding="utf-8")
+    blocks = [b for b in text.split("\n\n") if b.strip()]
+    ids = []
+    eos = sp.eos_id()
+    for b in blocks:
+        for line in b.splitlines():
+            if not line.strip(): continue
+            ids.extend(sp.encode(line, out_type=int)); ids.append(eos)
+        ids.append(eos)
+    ids = np.array(ids, dtype=np.int32)
+    n = len(ids); cut = int(n * 0.97)
+    train_ids = torch.tensor(ids[:cut], dtype=torch.long, device=DEVICE)
+    val_ids   = torch.tensor(ids[cut:], dtype=torch.long, device=DEVICE)
+    print(f"     tokens: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={sp.vocab_size()}")
+    return train_ids, val_ids, sp.vocab_size()
+
+# --------------------------
+# Tiny GPT model
+# --------------------------
+class CausalSelfAttention(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        assert n_embd % n_head == 0
+        self.n_head = n_head
+        self.head_dim = n_embd // n_head
+        self.qkv = nn.Linear(n_embd, 3*n_embd, bias=False)
+        self.proj = nn.Linear(n_embd, n_embd, bias=False)
+        self.attn_drop = nn.Dropout(dropout)
+        self.resid_drop = nn.Dropout(dropout)
+        self.register_buffer("mask", torch.tril(torch.ones(block_size, block_size)).view(1,1,block_size,block_size))
+    def forward(self, x):
+        B,T,C = x.shape
+        qkv = self.qkv(x); q,k,v = qkv.chunk(3, dim=-1)
+        q = q.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        k = k.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        v = v.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        att = (q @ k.transpose(-2,-1)) / math.sqrt(self.head_dim)
+        att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
+        att = torch.softmax(att, dim=-1)
+        att = self.attn_drop(att)
+        y = att @ v
+        y = y.transpose(1,2).contiguous().view(B,T,C)
+        y = self.resid_drop(self.proj(y))
+        return y
+
+class Block(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(n_embd)
+        self.attn = CausalSelfAttention(n_embd, n_head, dropout, block_size)
+        self.ln2 = nn.LayerNorm(n_embd)
+        self.mlp = nn.Sequential(
+            nn.Linear(n_embd, 4*n_embd),
+            nn.GELU(),
+            nn.Linear(4*n_embd, n_embd),
+            nn.Dropout(dropout),
+        )
+    def forward(self, x):
+        x = x + self.attn(self.ln1(x))
+        x = x + self.mlp(self.ln2(x))
+        return x
+
+class TinyGPT(nn.Module):
+    def __init__(self, vocab_size, n_layer, n_head, n_embd, block_size, dropout=0.0):
+        super().__init__()
+        self.block_size = block_size
+        self.tok_emb = nn.Embedding(vocab_size, n_embd)
+        self.pos_emb = nn.Embedding(block_size, n_embd)
+        self.blocks = nn.ModuleList([Block(n_embd, n_head, dropout, block_size) for _ in range(n_layer)])
+        self.ln_f = nn.LayerNorm(n_embd)
+        self.head = nn.Linear(n_embd, vocab_size, bias=False)
+        self.apply(self._init)
+    def _init(self, m):
+        if isinstance(m, (nn.Linear, nn.Embedding)): nn.init.normal_(m.weight, mean=0.0, std=0.02)
+        if isinstance(m, nn.Linear) and m.bias is not None: nn.init.zeros_(m.bias)
+        if isinstance(m, nn.LayerNorm): nn.init.ones_(m.weight); nn.init.zeros_(m.bias)
+    def forward(self, idx, targets=None):
+        B,T = idx.shape; assert T <= self.block_size
+        pos = torch.arange(0, T, device=idx.device)
+        x = self.tok_emb(idx) + self.pos_emb(pos)[None,:,:]
+        for blk in self.blocks: x = blk(x)
+        x = self.ln_f(x)
+        logits = self.head(x)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+    @torch.no_grad()
+    def generate(self, idx, max_new_tokens=200, temperature=0.8, top_k=50, top_p=0.95, repetition_penalty=1.0):
+        self.eval()
+        for _ in range(max_new_tokens):
+            idx_cond = idx[:, -self.block_size:]
+            logits, _ = self.forward(idx_cond)
+            logits = logits[:, -1, :]
+            if repetition_penalty != 1.0:
+                uniq, _ = torch.unique(idx_cond[0], return_counts=True)
+                logits[:, uniq] /= repetition_penalty
+            logits = logits / max(1e-8, temperature)
+            if top_k is not None:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                cutoff = v[:, -1].unsqueeze(-1)
+                logits = torch.where(logits < cutoff, torch.full_like(logits, -1e9), logits)
+            if top_p is not None:
+                sorted_logits, sorted_idx = torch.sort(logits, descending=True)
+                probs = torch.softmax(sorted_logits, dim=-1)
+                cdf = torch.cumsum(probs, dim=-1)
+                mask = cdf > top_p; mask[:, 0] = False
+                sorted_logits[mask] = -1e9
+                logits = torch.zeros_like(logits).scatter(1, sorted_idx, sorted_logits)
+            probs = torch.softmax(logits, dim=-1)
+            next_id = torch.multinomial(probs, num_samples=1)
+            idx = torch.cat([idx, next_id], dim=1)
+        return idx
+
+# --------------------------
+# Data loader for token IDs
+# --------------------------
+def get_batch(split_ids: torch.Tensor, B: int, T: int):
+    ix = torch.randint(0, split_ids.numel() - T - 1, (B,), device=split_ids.device)
+    x = torch.stack([split_ids[i:i+T] for i in ix])
+    y = torch.stack([split_ids[i+1:i+T+1] for i in ix])
+    return x, y
+
+# --------------------------
+# Train loop
+# --------------------------
+def train_model(vocab_size, train_ids, val_ids):
+    print("[5/6] Training tiny GPT on", DEVICE.type.upper(), "…")
+    model = TinyGPT(vocab_size, n_layer, n_head, n_embd, block_size, dropout).to(DEVICE)
+    params_m = sum(p.numel() for p in model.parameters())/1e6
+    print(f"     params: {params_m:.2f}M")
+    optimizer = torch.optim.AdamW(model.parameters(), lr=base_lr, betas=(0.9, 0.95), weight_decay=0.0)
+
+    use_amp = DEVICE.type == "cuda"
+    scaler = torch.amp.GradScaler("cuda", enabled=use_amp)
+    autocast = (lambda: torch.amp.autocast("cuda", dtype=torch.float16)) if use_amp else nullcontext
+
+    start = time.time()
+    best_val = float("inf")
+
+    def get_lr(step):
+        warmup = max(1, int(train_steps * warmup_ratio))
+        if step < warmup: return base_lr * (step+1)/warmup
+        progress = (step - warmup) / max(1, train_steps - warmup)
+        return min_lr + 0.5*(base_lr - min_lr)*(1 + math.cos(math.pi * min(1.0, progress)))
+
+    @torch.no_grad()
+    def eval_loss(iters=80):
+        model.eval(); losses=[]
+        for _ in range(iters):
+            xb, yb = get_batch(val_ids, min(batch_size, 32), block_size)
+            with autocast():
+                _, loss = model(xb, yb)
+            losses.append(loss.item())
+        model.train()
+        return float(sum(losses)/len(losses))
+
+    model.train(); step = 0
+    pbar = tqdm(total=train_steps, ncols=100, desc="[train]")
+    while step < train_steps and (time.time()-start) < MAX_SECONDS:
+        lr = get_lr(step)
+        for pg in optimizer.param_groups: pg["lr"] = lr
+        optimizer.zero_grad(set_to_none=True)
+
+        total_loss = 0.0
+        for _ in range(accum_steps):
+            xb, yb = get_batch(train_ids, batch_size, block_size)
+            with autocast():
+                _, loss = model(xb, yb)
+            if use_amp:
+                scaler.scale(loss / accum_steps).backward()
+            else:
+                (loss / accum_steps).backward()
+            total_loss += loss.item()
+
+        if use_amp:
+            scaler.unscale_(optimizer)
+        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+        if use_amp:
+            scaler.step(optimizer); scaler.update()
+        else:
+            optimizer.step()
+
+        step += 1; pbar.update(1)
+        if step % log_interval == 0 or step == 1:
+            pbar.set_postfix(train=f"{total_loss/accum_steps:.3f}", lr=f"{lr:.2e}")
+
+        if step % eval_every == 0:
+            vl = eval_loss()
+            best_val = min(best_val, vl)
+            print(f"     eval loss {vl:.3f} | best {best_val:.3f}")
+
+    pbar.close()
+    elapsed = time.time() - start
+    print(f"     done in {elapsed:.1f}s | best val {best_val:.3f}")
+
+    # Save model + config
+    ckpt_path = SAVE_DIR / "tinygpt.pt"
+    torch.save(model.state_dict(), ckpt_path)
+    (SAVE_DIR / "model_config.json").write_text(json.dumps({
+        "vocab_size": int(vocab_size),
+        "n_layer": n_layer, "n_head": n_head, "n_embd": n_embd,
+        "block_size": block_size, "dropout": dropout
+    }, indent=2))
+    print(f"[saved] weights → {ckpt_path}")
+    return model
+
+# --------------------------
+# Sampling helper
+# --------------------------
+def sample_chat(sp: spm.SentencePieceProcessor, model: TinyGPT, prompt: str, max_new_tokens=200):
+    prefix = f"You: {prompt}\nBot:"
+    ids = sp.encode(prefix, out_type=int)
+    x = torch.tensor([ids], dtype=torch.long, device=DEVICE)
+    with torch.no_grad():
+        y = model.generate(x, max_new_tokens=max_new_tokens, temperature=TEMP, top_k=TOP_K, top_p=TOP_P, repetition_penalty=REP_PEN)
+    return sp.decode(y[0].tolist())
+
+# --------------------------
+# Main
+# --------------------------
+def main():
+    # Build or reuse corpus/tokenizer/ids
+    corpus_path = SAVE_DIR / "corpus.txt"
+    spm_model   = SAVE_DIR / "spm_chat.model"
+    if not corpus_path.exists():
+        corpus_path = build_corpus()
+    else:
+        print(f"[cache] using {corpus_path}")
+
+    sp = spm.SentencePieceProcessor()
+    if not spm_model.exists():
+        sp = train_spm(corpus_path)
+    else:
+        sp.load(str(spm_model))
+        print(f"[cache] using {spm_model}")
+
+    enc_train = SAVE_DIR / "train_ids.pt"
+    enc_val   = SAVE_DIR / "val_ids.pt"
+    vocab_txt = SAVE_DIR / "vocab_size.txt"
+
+    if enc_train.exists() and enc_val.exists() and vocab_txt.exists():
+        train_ids = torch.load(enc_train, map_location=DEVICE)
+        val_ids   = torch.load(enc_val,   map_location=DEVICE)
+        vocab_size = int(vocab_txt.read_text())
+        print(f"[cache] loaded ids: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={vocab_size}")
+    else:
+        train_ids, val_ids, vocab_size = encode_corpus_to_ids(sp, corpus_path)
+        torch.save(train_ids, enc_train); torch.save(val_ids, enc_val)
+        vocab_txt.write_text(str(vocab_size))
+        print("[cache] saved encoded ids")
+
+    model = train_model(vocab_size, train_ids, val_ids)
+
+    print("\n[6/6] Samples:\n")
+    prompts = [
+        "Give me a spicy take on AI.",
+        "Roast my messy desk.",
+        "Explain recursion like you're annoyed.",
+        "Write a satirical headline about coffee.",
+        "Give me a shower thought about umbrellas.",
+        "Tell me a one-liner about deadlines.",
+        "Stay in Shakespeare mode and flatter me.",
+        "Reply sarcastically to: I love meetings.",
+    ]
+    out_path = SAVE_DIR / "samples.txt"
+    with out_path.open("w", encoding="utf-8") as f:
+        for p in prompts:
+            txt = sample_chat(sp, model, p, max_new_tokens=200)
+            print("----\n" + txt)
+            f.write("----\n" + txt + "\n")
+    print(f"\n[saved] samples → {out_path}")
+
+if __name__ == "__main__":
+    main()

train2.py

@@ -0,0 +1,1131 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+chat_sprint_standalone.py
+One-file pipeline: collect datasets -> reformat as You:/Bot: -> train tiny GPT (CUDA) -> sample & save model
+
+Requirements (install once):
+  uv pip install torch datasets sentencepiece tqdm numpy
+
+Run:
+  python chat_sprint_standalone.py
+"""
+
+import os, re, time, math, random, json
+from pathlib import Path
+from typing import List, Optional, Tuple
+from itertools import islice
+from contextlib import nullcontext
+
+import numpy as np
+from tqdm import tqdm
+from datasets import load_dataset, get_dataset_config_names
+import sentencepiece as spm
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# --------------------------
+# Global config (tweak here)
+# --------------------------
+SEED = 1337
+random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+if DEVICE.type == "cuda":
+    torch.set_float32_matmul_precision("high")
+    torch.backends.cuda.matmul.allow_tf32 = True
+
+SAVE_DIR = Path("./chat_sprint_artifacts")
+SAVE_DIR.mkdir(parents=True, exist_ok=True)
+
+# How many formatted You:/Bot: pairs to KEEP from each dataset (first run)
+CAPS = {
+    # NEW multi-turn / assistant cadence
+    "oasst1":      0,
+    "hhrlhf": 12000,
+    "ultrachat":   20000,
+    "dailydialog":  8000,
+    "bst":          6000,
+    "personachat":  8000,
+    "soda":        50000,
+    "topical":      8000,
+
+    # Existing single-turn & flavor
+    "shakespeare": 15000,
+    "jokes":       20000,
+    "dadjokes":     8000,
+    "rsarcasm":     8000,   # Thewillonline/reddit-sarcasm
+    "figlang":     12000,
+    "shower":       4000,   # HuggingFaceGECLM/REDDIT_submissions split "Showerthoughts"
+    "personas":     2000,
+    "tweeteval":    4000,
+    "fourchan":      500,   # dataset is tiny (~195)
+    "elonvtrump":   3000,
+}
+
+# Upper bound on rows to SCAN in each streaming dataset
+SCAN = {
+    # NEW
+    "oasst1":      200_000,
+    "hhrlhf": 100_000,
+    "ultrachat":   220_000,
+    "dailydialog":  30_000,
+    "bst":          30_000,
+    "personachat":  60_000,
+    "soda":        300_000,
+    "topical":      15_000,
+
+    # Existing
+    "jokes":       120_000,
+    "dadjokes":     60_000,
+    "rsarcasm":    120_000,
+    "figlang":     150_000,
+    "shower":      250_000,
+    "personas":     30_000,
+    "tweeteval":    60_000,
+    "fourchan":      2_000,
+    "elonvtrump":   60_000,
+}
+
+MAX_TOTAL_PAIRS = 150_000
+MAX_LEN = 120
+
+# Tokenizer
+VOCAB_SIZE = 5000
+TOKENIZER_PREFIX = SAVE_DIR / "spm_chat"
+USER_SYMBOLS = ["You:", "Bot:", "[STYLE=Snark]", "[FORM=TWEET]", "[FORM=HEADLINE]", "[MOOD=Unhinged]", "[MOOD=Cheeky]"]
+
+# Model size & train budget (~5 minutes on RTX 3090 with default)
+block_size   = 256
+n_layer      = 6
+n_head       = 6
+n_embd       = 384   # ~11.9M params
+dropout      = 0.0
+
+MAX_SECONDS  = 300   # hard cap
+train_steps  = 5000  # big number; time cap will stop near 5 min
+log_interval = 100
+eval_every   = 400
+batch_size   = 24
+accum_steps  = 3
+base_lr      = 3e-3
+min_lr       = 5e-4
+warmup_ratio = 0.06
+
+# Sampling defaults
+TEMP = 0.8
+TOP_K = 60
+TOP_P = 0.95
+REP_PEN = 1.08
+
+# --------------------------
+# Helpers: cleaning & format
+# --------------------------
+URL_RE   = re.compile(r"https?://\S+|www\.\S+", re.IGNORECASE)
+MENT_RE  = re.compile(r"@\w+")
+WS_RE    = re.compile(r"\s+")
+QUOTE_RE = re.compile(r"^[\"'“”‘’]+|[\"'“”‘’]+$")
+
+def clean_text(s: str) -> str:
+    s = s.strip()
+    s = URL_RE.sub("", s)
+    s = MENT_RE.sub("", s)
+    s = QUOTE_RE.sub("", s)
+    s = WS_RE.sub(" ", s)
+    return s.strip()
+
+def shorten_to(s: str, n: int) -> str:
+    s = re.sub(r"\s+", " ", s.strip())
+    if len(s) <= n: return s
+    cut = max(s.rfind(". ", 0, n), s.rfind("! ", 0, n), s.rfind("? ", 0, n))
+    if cut != -1: return s[:cut+1].strip()
+    return s[:n].strip()
+
+def keep_or_clip(s: str, min_len: int = 6, max_len: int = MAX_LEN) -> Optional[str]:
+    if not s: return None
+    s = re.sub(r"\s+", " ", s.strip())
+    if len(s) < min_len: return None
+    return shorten_to(s, max_len)
+
+def turn(you: str, bot: str, tags: str = "") -> str:
+    lines = [f"You: {you}".rstrip()]
+    if tags: lines.append(tags)
+    lines.append(f"Bot: {bot}".rstrip())
+    lines.append("")
+    return "\n".join(lines)
+
+def limited(ds, limit: int):
+    try:
+        return ds.take(limit)
+    except Exception:
+        return islice(ds, limit)
+
+def get_first_nonempty(ex, keys) -> Optional[str]:
+    for k in keys:
+        v = ex.get(k)
+        if isinstance(v, str) and v.strip():
+            return v
+    return None
+
+def to_str(x) -> Optional[str]:
+    if isinstance(x, str): return x
+    if isinstance(x, dict):
+        for k in ("text","utterance","content","response","value","message","msg"):
+            v = x.get(k)
+            if isinstance(v, str) and v.strip():
+                return v
+    return None
+
+# --------------------------
+# NEW Collectors (multi-turn)
+# --------------------------
+def collect_oasst1(pairs: List[str], overall: tqdm):
+    """OpenAssistant/oasst1: pair prompter -> assistant using parent links."""
+    try:
+        ds = load_dataset("OpenAssistant/oasst1", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["oasst1"], SCAN["oasst1"]
+        pbar = tqdm(total=keep_cap, desc="[oasst1]", unit="pair", leave=False, ncols=100)
+        kept = scanned = 0
+        seen = {}  # id -> (role, text)
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            role = str(ex.get("role") or "")
+            txt = keep_or_clip(clean_text(str(ex.get("text") or "")))
+            mid = ex.get("message_id") or ex.get("id")
+            pid = ex.get("parent_id")
+            if role == "assistant" and pid in seen and seen[pid][0] == "prompter":
+                you = seen[pid][1]
+                if you and txt:
+                    pairs.append(turn(you, txt))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= keep_cap: break
+            if mid and role in ("assistant", "prompter"):
+                seen[mid] = (role, txt or "")
+        pbar.close()
+        print(f"[ok] oasst1 kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] oasst1: {e}")
+
+def collect_ultrachat(pairs: List[str], overall: tqdm):
+    """HuggingFaceH4/ultrachat_200k: slide over user->assistant pairs."""
+    try:
+        ds = load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft", streaming=True)
+        keep_cap, scan_cap = CAPS["ultrachat"], SCAN["ultrachat"]
+        pbar = tqdm(total=keep_cap, desc="[ultrachat]", unit="pair", leave=False, ncols=100)
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            msgs = ex.get("messages") or []
+            for a, b in zip(msgs, msgs[1:]):
+                if (a.get("role") in ("user","human")) and (b.get("role") in ("assistant","gpt")):
+                    you = keep_or_clip(clean_text(str(a.get("content") or "")))
+                    bot = keep_or_clip(clean_text(str(b.get("content") or "")))
+                    if you and bot:
+                        pairs.append(turn(you, bot))
+                        kept += 1; pbar.update(1); overall.update(1)
+                        if kept >= keep_cap: break
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] ultrachat kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] ultrachat: {e}")
+
+
+def collect_dailydialog(pairs: List[str], overall: tqdm):
+    """
+    DailyDialog via data-only mirrors (no dataset scripts).
+    Tries Parquet-converted branches/configs first; collects adjacent utterance pairs.
+    """
+    keep_cap, scan_cap = CAPS["dailydialog"], SCAN["dailydialog"]
+    candidates = [
+        # (dataset_id, name_or_None, revision_or_None)
+        ("roskoN/dailydialog", "full",    "refs/convert/parquet"),   # has full/train/*.parquet
+        ("roskoN/dailydialog", "default", "refs/convert/parquet"),
+        ("frankdarkluo/DailyDialog", "default", "refs/convert/parquet"),
+        ("ConvLab/dailydialog", None, None),  # sometimes ships plain JSON/arrow
+    ]
+
+    kept = 0
+    for dsid, name, rev in candidates:
+        try:
+            # Build load_dataset call dynamically
+            load_kwargs = dict(split="train", streaming=True)
+            if rev is not None:
+                load_kwargs["revision"] = rev
+            if name is not None:
+                ds = load_dataset(dsid, name, **load_kwargs)
+            else:
+                ds = load_dataset(dsid, **load_kwargs)
+
+            pbar = tqdm(total=keep_cap, desc=f"[dailydialog:{dsid}]", unit="pair", leave=False, ncols=100)
+            for ex in limited(ds, scan_cap):
+                # Flexible extraction across mirrors
+                dialog = (
+                    ex.get("dialog") or ex.get("dialogue") or ex.get("utterances")
+                    or ex.get("turns") or ex.get("content") or []
+                )
+                seq = []
+                if isinstance(dialog, list):
+                    for u in dialog:
+                        s = to_str(u) if not isinstance(u, str) else u
+                        s = keep_or_clip(clean_text(str(s or "")))
+                        if s:
+                            seq.append(s)
+
+                for a, b in zip(seq, seq[1:]):
+                    pairs.append(turn(a, b))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= keep_cap:
+                        break
+                if kept >= keep_cap:
+                    break
+            pbar.close()
+
+            if kept > 0:
+                print(f"[ok] dailydialog kept={kept} via {dsid} ({'name='+name if name else 'no-name'} {rev or ''})")
+                return
+            else:
+                print(f"[try next] dailydialog: {dsid} produced 0 pairs; trying next candidate…")
+        except Exception as e:
+            print(f"[try next] dailydialog {dsid}: {e}")
+
+    print("[skip] dailydialog: no usable Parquet/JSON mirror found (all candidates failed)")
+
+
+def collect_bst(pairs: List[str], overall: tqdm):
+    """blended_skill_talk: previous_utterance -> one of guided/free/suggested."""
+    try:
+        ds = load_dataset("blended_skill_talk", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["bst"], SCAN["bst"]
+        pbar = tqdm(total=keep_cap, desc="[bst]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        def pick_first(x):
+            if isinstance(x, list) and x: return x[0]
+            if isinstance(x, dict):
+                for v in x.values():
+                    if isinstance(v, list) and v: return v[0]
+            return None
+        for ex in limited(ds, scan_cap):
+            you = keep_or_clip(clean_text(str(ex.get("previous_utterance") or ex.get("context") or "")))
+            cand = pick_first(ex.get("guided_messages")) or pick_first(ex.get("free_messages")) or pick_first(ex.get("suggestions"))
+            bot = keep_or_clip(clean_text(str(cand or "")))
+            if you and bot:
+                pairs.append(turn(you, bot))
+                kept += 1; pbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] bst kept={kept}")
+    except Exception as e:
+        print(f"[skip] bst: {e}")
+
+def collect_personachat(pairs: List[str], overall: tqdm):
+    """bavard/personachat_truecased (Parquet): expand history[] into adjacent pairs."""
+    try:
+        ds = load_dataset("bavard/personachat_truecased", split="train", streaming=True, revision="refs/convert/parquet")
+        keep_cap, scan_cap = CAPS["personachat"], SCAN["personachat"]
+        pbar = tqdm(total=keep_cap, desc="[personachat]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in limited(ds, scan_cap):
+            hist = ex.get("history") or []
+            seq = [keep_or_clip(clean_text(str(u))) for u in hist]
+            seq = [s for s in seq if s]
+            for a, b in zip(seq, seq[1:]):
+                pairs.append(turn(a, b))
+                kept += 1; pbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] personachat kept={kept}")
+    except Exception as e:
+        print(f"[skip] personachat: {e}")
+
+
+def collect_soda(pairs: List[str], overall: tqdm):
+    """allenai/soda: dialogue can be list[str] or list[dict]."""
+    try:
+        ds = load_dataset("allenai/soda", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["soda"], SCAN["soda"]
+        pbar = tqdm(total=keep_cap, desc="[soda]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in limited(ds, scan_cap):
+            dia = ex.get("dialogue") or ex.get("dialog") or ex.get("utterances") or []
+            seq = []
+            for u in dia:
+                s = to_str(u)
+                s = keep_or_clip(clean_text(str(s or "")))
+                if s: seq.append(s)
+            for a, b in zip(seq, seq[1:]):
+                pairs.append(turn(a, b))
+                kept += 1; pbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] soda kept={kept}")
+    except Exception as e:
+        print(f"[skip] soda: {e}")
+
+def collect_topical_chat(pairs: List[str], overall: tqdm):
+    """
+    Topical-Chat: try Hub mirrors, else fetch official GitHub JSON files.
+    Keeps only conversation text (no reading sets).
+    """
+    keep_cap, scan_cap = CAPS["topical"], SCAN["topical"]
+    kept = 0
+
+    # 1) Try Hub mirrors/splits first
+    dsids = [
+        "Conversational-Reasoning/Topical-Chat",
+        "AmazonScience/Topical-Chat",
+        "microsoft/Topical-Chat",
+    ]
+    splits = ["train","valid_freq","valid_rare","test_freq","test_rare","validation","test","valid_frequent","test_frequent"]
+
+    for dsid in dsids:
+        try:
+            pbar = tqdm(total=keep_cap, desc=f"[topical:{dsid}]", unit="pair", leave=False, ncols=100)
+            for split in splits:
+                try:
+                    ds = load_dataset(dsid, split=split, streaming=True)
+                except Exception:
+                    continue
+                for ex in limited(ds, scan_cap):
+                    dia = ex.get("messages") or ex.get("conversation") or ex.get("utterances") or ex.get("dialogue") or ex.get("content") or []
+                    seq = []
+                    for u in dia:
+                        s = to_str(u)
+                        s = keep_or_clip(clean_text(str(s or "")))
+                        if s: seq.append(s)
+                    for a, b in zip(seq, seq[1:]):
+                        pairs.append(turn(a, b))
+                        kept += 1; pbar.update(1); overall.update(1)
+                        if kept >= keep_cap: break
+                    if kept >= keep_cap: break
+                if kept >= keep_cap: break
+            pbar.close()
+            if kept > 0:
+                print(f"[ok] topical_chat ({dsid}) kept={kept}")
+                return
+        except Exception as e:
+            print(f"[try next] topical_chat {dsid}: {e}")
+
+    # 2) Fallback to official GitHub JSON (robust)
+    try:
+        names = ["train","valid_freq","valid_rare","test_freq","test_rare"]
+        base = "https://raw.githubusercontent.com/alexa/Topical-Chat/master/conversations"
+        pbar = tqdm(total=keep_cap, desc="[topical:github]", unit="pair", leave=False, ncols=100)
+        for nm in names:
+            try:
+                with urlreq.urlopen(f"{base}/{nm}.json") as r:
+                    data = json.loads(r.read().decode("utf-8"))
+            except Exception:
+                continue
+            for _, convo in data.items():
+                content = convo.get("content") or []
+                seq = []
+                for t in content:
+                    s = keep_or_clip(clean_text(str(t.get("message") or t.get("text") or "")))
+                    if s: seq.append(s)
+                for a, b in zip(seq, seq[1:]):
+                    pairs.append(turn(a, b))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= keep_cap: break
+                if kept >= keep_cap: break
+            if kept >= keep_cap: break
+        pbar.close()
+        if kept == 0:
+            print("[skip] topical_chat: no usable conversations found")
+        else:
+            print(f"[ok] topical_chat (github) kept={kept}")
+    except Exception as e:
+        print(f"[skip] topical_chat (github): {e}")
+
+
+# --------------------------
+# Collectors (existing)
+# --------------------------
+def collect_shakespeare(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset(
+            "text",
+            data_files={"train": "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"},
+            split="train",
+            streaming=True,
+        )
+        kept = 0
+        pbar = tqdm(total=CAPS["shakespeare"], desc="[shakespeare]", unit="pair", leave=False, ncols=100)
+        for row in ds:
+            txt = keep_or_clip(clean_text(row["text"]))
+            if not txt: continue
+            pairs.append(turn("Continue in Shakespearean style.", txt))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= CAPS["shakespeare"]: break
+        pbar.close()
+        print(f"[ok] shakespeare kept={kept}")
+    except Exception as e:
+        print(f"[skip] shakespeare: {e}")
+
+def collect_reddit_jokes(pairs: List[str], overall: tqdm):
+    for dsid in ["SocialGrep/one-million-reddit-jokes", "SocialGrep/reddit_jokes", "timc1/reddit_jokes"]:
+        try:
+            ds = load_dataset(dsid, split="train", streaming=True)
+            kept = 0
+            pbar = tqdm(total=CAPS["jokes"], desc="[jokes]", unit="pair", leave=False, ncols=100)
+            for ex in limited(ds, SCAN["jokes"]):
+                title = keep_or_clip(clean_text(str(ex.get("title") or "")))
+                body  = keep_or_clip(clean_text(str(ex.get("selftext") or ex.get("body") or "")))
+                if body and title:
+                    pairs.append(turn(title, body))
+                elif title:
+                    pairs.append(turn("Tell me a short joke.", title))
+                else:
+                    continue
+                kept += 1; pbar.update(1); overall.update(1)
+                if kept >= CAPS["jokes"]: break
+            pbar.close()
+            print(f"[ok] jokes {dsid} kept={kept}")
+            return
+        except Exception as e:
+            print(f"[try next] jokes {dsid}: {e}")
+    print("[skip] jokes: none worked")
+
+def collect_dadjokes(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("shuttie/reddit-dadjokes", split="train", streaming=True)
+        kept = 0
+        pbar = tqdm(total=CAPS["dadjokes"], desc="[dadjokes]", unit="pair", leave=False, ncols=100)
+        for ex in limited(ds, SCAN["dadjokes"]):
+            setup = keep_or_clip(clean_text(str(ex.get("setup") or ex.get("instruction") or ex.get("input") or "")))
+            punch = keep_or_clip(clean_text(str(ex.get("punchline") or ex.get("output") or "")))
+            if not (setup and punch): continue
+            pairs.append(turn(setup, punch))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= CAPS["dadjokes"]: break
+        pbar.close()
+        print(f"[ok] dadjokes kept={kept}")
+    except Exception as e:
+        print(f"[skip] dadjokes: {e}")
+
+def collect_reddit_sarcasm(pairs: List[str], overall: tqdm):
+    """Thewillonline/reddit-sarcasm — flexible parsing, scan+keep bars."""
+    try:
+        ds = load_dataset("Thewillonline/reddit-sarcasm", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["rsarcasm"], SCAN["rsarcasm"]
+        scanbar = tqdm(total=scan_cap, desc="[sarcasm scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[sarcasm kept]", unit="pair", leave=False, ncols=100)
+
+        PATS = [
+            re.compile(r"User\s*:\s*(.+?)\s*(?:Reddit\s*Comment|Comment|Reply)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
+            re.compile(r"Post\s*:\s*(.+?)\s*(?:Top\s*Comment|Comment)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
+        ]
+        def parse(raw: str) -> Tuple[Optional[str], Optional[str]]:
+            raw = raw.replace("<|endoftext|>", "\n")
+            for pat in PATS:
+                m = pat.search(raw)
+                if m:
+                    return m.group(1).strip(), m.group(2).strip()
+            lines = [ln.strip() for ln in raw.splitlines() if ln.strip()]
+            if len(lines) >= 2: return lines[0], lines[1]
+            if len(lines) == 1: return "Reply with sarcasm:", lines[0]
+            return None, None
+
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            you, bot = parse(str(ex.get("text") or ""))
+            you = keep_or_clip(you); bot = keep_or_clip(bot)
+            if you and bot:
+                pairs.append(turn(you, bot, "[STYLE=Snark]"))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+            if scanned % 2000 == 0:
+                keepbar.set_postfix(rate=f"{kept/max(1,scanned):.2%}")
+        scanbar.close(); keepbar.close()
+        print(f"[ok] reddit-sarcasm kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] reddit-sarcasm: {e}")
+
+def collect_figlang(pairs: List[str], overall: tqdm):
+    for dsid in ["tasksource/figlang2020-sarcasm", "tasksource/figlang2020_sarcasm"]:
+        try:
+            ds = load_dataset(dsid, split="train", streaming=True)
+            kept = 0
+            pbar = tqdm(total=CAPS["figlang"], desc="[figlang]", unit="pair", leave=False, ncols=100)
+            for ex in limited(ds, SCAN["figlang"]):
+                ctx = ex.get("context")
+                if isinstance(ctx, list) and ctx:
+                    context_str = " ".join(str(c) for c in ctx[-2:])
+                else:
+                    context_str = str(ex.get("context") or ex.get("prompt") or "")
+                reply = str(ex.get("response") or ex.get("answer") or ex.get("text") or "")
+                context_str = keep_or_clip(clean_text(context_str))
+                reply      = keep_or_clip(clean_text(reply))
+                if reply:
+                    if context_str:
+                        pairs.append(turn(context_str, reply, "[STYLE=Snark]"))
+                    else:
+                        pairs.append(turn("Reply with sarcasm:", reply, "[STYLE=Snark]"))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= CAPS["figlang"]: break
+            pbar.close()
+            print(f"[ok] figlang {dsid} kept={kept}")
+            return
+        except Exception as e:
+            print(f"[try next] figlang {dsid}: {e}")
+    print("[skip] figlang")
+
+def collect_showerthoughts(pairs: List[str], overall: tqdm):
+    """Use REEDIT_submissions split 'Showerthoughts' directly (no 'train')."""
+    try:
+        ds = load_dataset("HuggingFaceGECLM/REDDIT_submissions", split="Showerthoughts", streaming=True)
+        keep_cap, scan_cap = CAPS["shower"], SCAN["shower"]
+        scanbar = tqdm(total=scan_cap, desc="[shower scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[shower kept]", unit="pair", leave=False, ncols=100)
+
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            title = get_first_nonempty(ex, ["title", "selftext", "text"]) or ""
+            text = keep_or_clip(clean_text(title))
+            if text:
+                pairs.append(turn("Give me a shower thought.", text))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+        scanbar.close(); keepbar.close()
+        print(f"[ok] showerthoughts kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] showerthoughts: {e}")
+
+def collect_personas(pairs: List[str], overall: tqdm):
+    """Non-streaming is more reliable for this dataset."""
+    try:
+        ds = load_dataset("NapthaAI/twitter_personas")["train"]
+        keep_cap = CAPS["personas"]
+        pbar = tqdm(total=keep_cap, desc="[personas]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in ds:
+            desc = get_first_nonempty(ex, ["description","persona","bio","text","content","full_text"])
+            if not isinstance(desc, str) and isinstance(ex.get("content"), dict):
+                desc = ex["content"].get("text")
+            desc = keep_or_clip(clean_text(str(desc or "")))
+            if not desc: continue
+            pairs.append(turn("Adopt this persona in one sentence.", desc, "[FORM=TWEET]"))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] personas kept={kept}")
+    except Exception as e:
+        print(f"[skip] personas: {e}")
+
+def collect_tweeteval(pairs: List[str], overall: tqdm):
+    """Handle super_tweeteval (text_1/text_2, etc.) and fallback tweet_eval."""
+    def extract_pair(ex):
+        t = ex.get("text")
+        if isinstance(t, str) and t.strip():
+            return "React with a sharp one-liner.", t
+        for a,b in [("text_1","text_2"), ("sentence1","sentence2"),
+                    ("premise","hypothesis"), ("question","answer"),
+                    ("context","response"), ("tweet1","tweet2")]:
+            t1, t2 = ex.get(a), ex.get(b)
+            if isinstance(t1, str) and t1.strip() and isinstance(t2, str) and t2.strip():
+                return t1, t2
+        return None
+
+    def run_on(dsname, pick, is_super):
+        keep_cap, scan_cap = CAPS["tweeteval"], SCAN["tweeteval"]
+        pbar = tqdm(total=keep_cap, desc=f"[tweeteval:{pick}]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        ds = load_dataset(dsname, pick, split="train", streaming=True)
+        for ex in limited(ds, scan_cap):
+            pair = extract_pair(ex) if is_super else ("React with a sharp one-liner.", ex.get("text")) if ex.get("text") else None
+            if not pair: continue
+            you, bot = pair
+            you = keep_or_clip(clean_text(str(you or ""))); bot = keep_or_clip(clean_text(str(bot or "")))
+            if not (you and bot): continue
+            tag = "[STYLE=Snark]" if you and you != "React with a sharp one-liner." else ""
+            pairs.append(turn(you, bot, tag))
+            kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        return kept
+
+    kept_total = 0
+    try:
+        cfgs = get_dataset_config_names("cardiffnlp/super_tweeteval")
+    except Exception:
+        cfgs = []
+    prio = ["irony","sarcasm","humor","sentiment","emoji","emotion","stance","offensive","hate"]
+    ordered = [c for c in prio if c in cfgs] + [c for c in cfgs if c not in prio]
+    for pick in ordered:
+        kept_total += run_on("cardiffnlp/super_tweeteval", pick, True)
+        if kept_total >= CAPS["tweeteval"]:
+            print(f"[ok] tweeteval(super) kept={kept_total}"); return
+    if kept_total == 0:
+        try:
+            base_cfgs = get_dataset_config_names("cardiffnlp/tweet_eval")
+        except Exception:
+            base_cfgs = []
+        ordered_b = [c for c in prio if c in base_cfgs] + [c for c in base_cfgs if c not in prio]
+        for pick in ordered_b:
+            kept_total += run_on("cardiffnlp/tweet_eval", pick, False)
+            if kept_total >= CAPS["tweeteval"]:
+                print(f"[ok] tweeteval(base) kept={kept_total}"); return
+    print(f"[ok] tweeteval kept={kept_total}")
+
+def collect_fourchan(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("sbussiso/4chan-dataset", split="train", streaming=True)
+        keep_cap = min(CAPS["fourchan"], 195)
+        pbar = tqdm(total=keep_cap, desc="[4chan]", unit="pair", leave=False, ncols=100)
+        kept = 0
+        for ex in limited(ds, SCAN["fourchan"]):
+            prompt = keep_or_clip(clean_text(str(ex.get("prompt") or "")))
+            resp   = keep_or_clip(clean_text(str(ex.get("response") or "")))
+            if prompt and resp:
+                pairs.append(turn(prompt, resp))
+                kept += 1; pbar.update(1); overall.update(1)
+            else:
+                txt = keep_or_clip(clean_text(str(ex.get("text") or ex.get("body") or ex.get("content") or "")))
+                if txt:
+                    pairs.append(turn("Drop a spicy one-liner.", txt))
+                    kept += 1; pbar.update(1); overall.update(1)
+            if kept >= keep_cap: break
+        pbar.close()
+        print(f"[ok] 4chan kept={kept}")
+    except Exception as e:
+        print(f"[skip] 4chan: {e}")
+
+def collect_elon_trump(pairs: List[str], overall: tqdm):
+    try:
+        ds = load_dataset("MasaFoundation/Twitter_X_Elon_vs_Trump", split="train", streaming=True, revision="refs/convert/parquet")
+        keep_cap, scan_cap = CAPS["elonvtrump"], SCAN["elonvtrump"]
+        scanbar = tqdm(total=scan_cap, desc="[elon_vs_trump scan]", unit="row", leave=False, ncols=100)
+        keepbar = tqdm(total=keep_cap, desc="[elon_vs_trump kept]", unit="pair", leave=False, ncols=100)
+        kept = scanned = 0
+        for ex in limited(ds, scan_cap):
+            scanned += 1
+            txt = get_first_nonempty(ex, ["content","text","tweet","full_text"]) or ""
+            txt = keep_or_clip(clean_text(txt))
+            if txt:
+                pairs.append(turn("[FORM=TWEET] One sentence hot take:", txt, "[FORM=TWEET]"))
+                kept += 1; keepbar.update(1); overall.update(1)
+                if kept >= keep_cap: break
+            scanbar.update(1)
+        scanbar.close(); keepbar.close()
+        print(f"[ok] Elon_vs_Trump kept={kept} (scanned {scanned})")
+    except Exception as e:
+        print(f"[skip] Elon_vs_Trump: {e}")
+        
+        
+def collect_hh_rlhf(pairs: List[str], overall: tqdm):
+    """
+    Anthropic HH-RLHF (English multi-turn). We parse the 'chosen' conversation
+    into adjacent Human->Assistant pairs and emit You:/Bot: blocks.
+    """
+    try:
+        ds = load_dataset("Anthropic/hh-rlhf", split="train", streaming=True)
+        keep_cap, scan_cap = CAPS["hhrlhf"], SCAN["hhrlhf"]
+        pbar = tqdm(total=keep_cap, desc="[hh-rlhf]", unit="pair", leave=False, ncols=100)
+        kept = 0
+
+        HUMAN = re.compile(r"(?:^|\n)\s*Human:\s*", re.IGNORECASE)
+        ASSIST = re.compile(r"(?:^|\n)\s*Assistant:\s*", re.IGNORECASE)
+
+        for ex in limited(ds, scan_cap):
+            # Prefer the curated 'chosen' conversation; fall back to 'prompt' if needed
+            convo = ex.get("chosen") or ex.get("prompt") or ""
+            if not isinstance(convo, str) or not convo.strip():
+                continue
+
+            # Split into alternating roles based on explicit markers
+            # We'll rebuild a list of (role, text) turns.
+            tokens = re.split(r"(Human:|Assistant:)", convo)
+            turns = []
+            # tokens looks like: ["", "Human:", "text...", "Assistant:", "text...", ...]
+            for i in range(1, len(tokens), 2):
+                role = tokens[i].strip().lower()
+                content = tokens[i + 1] if i + 1 < len(tokens) else ""
+                content = keep_or_clip(clean_text(str(content or "")))
+                if not content:
+                    continue
+                if role.startswith("human"):
+                    turns.append(("user", content))
+                elif role.startswith("assistant"):
+                    turns.append(("assistant", content))
+
+            # Emit adjacent user->assistant pairs
+            for a, b in zip(turns, turns[1:]):
+                if a[0] == "user" and b[0] == "assistant":
+                    pairs.append(turn(a[1], b[1]))
+                    kept += 1; pbar.update(1); overall.update(1)
+                    if kept >= keep_cap:
+                        break
+            if kept >= keep_cap:
+                break
+
+        pbar.close()
+        print(f"[ok] hh-rlhf kept={kept}")
+    except Exception as e:
+        print(f"[skip] hh-rlhf: {e}")
+        
+
+# --------------------------
+# Build corpus
+# --------------------------
+def build_corpus() -> Path:
+    pairs: List[str] = []
+    total_target = sum(CAPS.values())
+    print("[1/6] Collecting & reformatting datasets (streaming, capped)…")
+    overall = tqdm(total=total_target, desc="[all] collecting", unit="pair", ncols=100)
+
+    collectors = [
+        collect_oasst1,
+        collect_hh_rlhf,
+        collect_ultrachat,
+        collect_dailydialog,
+        collect_bst,
+        collect_personachat,
+        collect_soda,
+        collect_topical_chat,   # optional; will skip if not available
+        collect_shakespeare,
+        collect_reddit_jokes,
+        collect_dadjokes,
+        collect_reddit_sarcasm,
+        collect_figlang,
+        collect_showerthoughts,
+        collect_personas,
+        collect_tweeteval,
+        collect_fourchan,
+        collect_elon_trump,
+    ]
+    for fn in collectors:
+        try:
+            fn(pairs, overall)
+        except Exception as e:
+            print(f"[collector error] {fn.__name__}: {e}")
+
+    overall.close()
+    print("[2/6] Deduplicating & clipping…")
+    seen = set(); deduped = []
+    for block in pairs:
+        try:
+            bot_line = [ln for ln in block.splitlines() if ln.startswith("Bot:")][0]
+            key = bot_line[4:].strip().lower()
+        except Exception:
+            key = block.strip().lower()
+        if key in seen: continue
+        seen.add(key); deduped.append(block)
+
+    random.shuffle(deduped)
+    if len(deduped) > MAX_TOTAL_PAIRS:
+        deduped = deduped[:MAX_TOTAL_PAIRS]
+
+    out_path = SAVE_DIR / "corpus.txt"
+    out_path.write_text("\n".join(deduped), encoding="utf-8")
+    print(f"     wrote {len(deduped)} pairs → {out_path}")
+    return out_path
+
+# --------------------------
+# SentencePiece tokenizer
+# --------------------------
+def train_spm(corpus_path: Path) -> spm.SentencePieceProcessor:
+    print("[3/6] Training SentencePiece tokenizer…")
+    spm.SentencePieceTrainer.Train(
+        input=str(corpus_path),
+        model_prefix=str(TOKENIZER_PREFIX),
+        vocab_size=VOCAB_SIZE,
+        model_type="unigram",
+        character_coverage=1.0,
+        user_defined_symbols=USER_SYMBOLS,
+        bos_id=1, eos_id=2, unk_id=0, pad_id=-1
+    )
+    sp = spm.SentencePieceProcessor()
+    sp.load(f"{TOKENIZER_PREFIX}.model")
+    print(f"     tokenizer saved at {TOKENIZER_PREFIX}.model")
+    return sp
+
+# --------------------------
+# Encode to token IDs
+# --------------------------
+def encode_corpus_to_ids(sp: spm.SentencePieceProcessor, corpus_path: Path):
+    print("[4/6] Encoding corpus to token IDs…")
+    text = corpus_path.read_text(encoding="utf-8")
+    blocks = [b for b in text.split("\n\n") if b.strip()]
+    ids = []
+    eos = sp.eos_id()
+    for b in blocks:
+        for line in b.splitlines():
+            if not line.strip(): continue
+            ids.extend(sp.encode(line, out_type=int)); ids.append(eos)
+        ids.append(eos)
+    ids = np.array(ids, dtype=np.int32)
+    n = len(ids); cut = int(n * 0.97)
+    train_ids = torch.tensor(ids[:cut], dtype=torch.long, device=DEVICE)
+    val_ids   = torch.tensor(ids[cut:], dtype=torch.long, device=DEVICE)
+    print(f"     tokens: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={sp.vocab_size()}")
+    return train_ids, val_ids, sp.vocab_size()
+
+# --------------------------
+# Tiny GPT model
+# --------------------------
+class CausalSelfAttention(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        assert n_embd % n_head == 0
+        self.n_head = n_head
+        self.head_dim = n_embd // n_head
+        self.qkv = nn.Linear(n_embd, 3*n_embd, bias=False)
+        self.proj = nn.Linear(n_embd, n_embd, bias=False)
+        self.attn_drop = nn.Dropout(dropout)
+        self.resid_drop = nn.Dropout(dropout)
+        self.register_buffer("mask", torch.tril(torch.ones(block_size, block_size)).view(1,1,block_size,block_size))
+    def forward(self, x):
+        B,T,C = x.shape
+        qkv = self.qkv(x); q,k,v = qkv.chunk(3, dim=-1)
+        q = q.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        k = k.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        v = v.view(B,T,self.n_head,self.head_dim).transpose(1,2)
+        att = (q @ k.transpose(-2,-1)) / math.sqrt(self.head_dim)
+        att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
+        att = torch.softmax(att, dim=-1)
+        att = self.attn_drop(att)
+        y = att @ v
+        y = y.transpose(1,2).contiguous().view(B,T,C)
+        y = self.resid_drop(self.proj(y))
+        return y
+
+class Block(nn.Module):
+    def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(n_embd)
+        self.attn = CausalSelfAttention(n_embd, n_head, dropout, block_size)
+        self.ln2 = nn.LayerNorm(n_embd)
+        self.mlp = nn.Sequential(
+            nn.Linear(n_embd, 4*n_embd),
+            nn.GELU(),
+            nn.Linear(4*n_embd, n_embd),
+            nn.Dropout(dropout),
+        )
+    def forward(self, x):
+        x = x + self.attn(self.ln1(x))
+        x = x + self.mlp(self.ln2(x))
+        return x
+
+class TinyGPT(nn.Module):
+    def __init__(self, vocab_size, n_layer, n_head, n_embd, block_size, dropout=0.0):
+        super().__init__()
+        self.block_size = block_size
+        self.tok_emb = nn.Embedding(vocab_size, n_embd)
+        self.pos_emb = nn.Embedding(block_size, n_embd)
+        self.blocks = nn.ModuleList([Block(n_embd, n_head, dropout, block_size) for _ in range(n_layer)])
+        self.ln_f = nn.LayerNorm(n_embd)
+        self.head = nn.Linear(n_embd, vocab_size, bias=False)
+        self.apply(self._init)
+    def _init(self, m):
+        if isinstance(m, (nn.Linear, nn.Embedding)): nn.init.normal_(m.weight, mean=0.0, std=0.02)
+        if isinstance(m, nn.Linear) and m.bias is not None: nn.init.zeros_(m.bias)
+        if isinstance(m, nn.LayerNorm): nn.init.ones_(m.weight); nn.init.zeros_(m.bias)
+    def forward(self, idx, targets=None):
+        B,T = idx.shape; assert T <= self.block_size
+        pos = torch.arange(0, T, device=idx.device)
+        x = self.tok_emb(idx) + self.pos_emb(pos)[None,:,:]
+        for blk in self.blocks: x = blk(x)
+        x = self.ln_f(x)
+        logits = self.head(x)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+    @torch.no_grad()
+    def generate(self, idx, max_new_tokens=200, temperature=0.8, top_k=50, top_p=0.95, repetition_penalty=1.0):
+        self.eval()
+        for _ in range(max_new_tokens):
+            idx_cond = idx[:, -self.block_size:]
+            logits, _ = self.forward(idx_cond)
+            logits = logits[:, -1, :]
+            if repetition_penalty != 1.0:
+                uniq, _ = torch.unique(idx_cond[0], return_counts=True)
+                logits[:, uniq] /= repetition_penalty
+            logits = logits / max(1e-8, temperature)
+            if top_k is not None:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                cutoff = v[:, -1].unsqueeze(-1)
+                logits = torch.where(logits < cutoff, torch.full_like(logits, -1e9), logits)
+            if top_p is not None:
+                sorted_logits, sorted_idx = torch.sort(logits, descending=True)
+                probs = torch.softmax(sorted_logits, dim=-1)
+                cdf = torch.cumsum(probs, dim=-1)
+                mask = cdf > top_p; mask[:, 0] = False
+                sorted_logits[mask] = -1e9
+                logits = torch.zeros_like(logits).scatter(1, sorted_idx, sorted_logits)
+            probs = torch.softmax(logits, dim=-1)
+            next_id = torch.multinomial(probs, num_samples=1)
+            idx = torch.cat([idx, next_id], dim=1)
+        return idx
+
+# --------------------------
+# Data loader for token IDs
+# --------------------------
+def get_batch(split_ids: torch.Tensor, B: int, T: int):
+    ix = torch.randint(0, split_ids.numel() - T - 1, (B,), device=split_ids.device)
+    x = torch.stack([split_ids[i:i+T] for i in ix])
+    y = torch.stack([split_ids[i+1:i+T+1] for i in ix])
+    return x, y
+
+# --------------------------
+# Train loop
+# --------------------------
+def train_model(vocab_size, train_ids, val_ids):
+    print("[5/6] Training tiny GPT on", DEVICE.type.upper(), "…")
+    model = TinyGPT(vocab_size, n_layer, n_head, n_embd, block_size, dropout).to(DEVICE)
+    params_m = sum(p.numel() for p in model.parameters())/1e6
+    print(f"     params: {params_m:.2f}M")
+    optimizer = torch.optim.AdamW(model.parameters(), lr=base_lr, betas=(0.9, 0.95), weight_decay=0.0)
+
+    use_amp = DEVICE.type == "cuda"
+    scaler = torch.amp.GradScaler("cuda", enabled=use_amp)
+    autocast = (lambda: torch.amp.autocast("cuda", dtype=torch.float16)) if use_amp else nullcontext
+
+    start = time.time()
+    best_val = float("inf")
+
+    def get_lr(step):
+        warmup = max(1, int(train_steps * warmup_ratio))
+        if step < warmup: return base_lr * (step+1)/warmup
+        progress = (step - warmup) / max(1, train_steps - warmup)
+        return min_lr + 0.5*(base_lr - min_lr)*(1 + math.cos(math.pi * min(1.0, progress)))
+
+    @torch.no_grad()
+    def eval_loss(iters=80):
+        model.eval(); losses=[]
+        for _ in range(iters):
+            xb, yb = get_batch(val_ids, min(batch_size, 32), block_size)
+            with autocast():
+                _, loss = model(xb, yb)
+            losses.append(loss.item())
+        model.train()
+        return float(sum(losses)/len(losses))
+
+    model.train(); step = 0
+    pbar = tqdm(total=train_steps, ncols=100, desc="[train]")
+    while step < train_steps and (time.time()-start) < MAX_SECONDS:
+        lr = get_lr(step)
+        for pg in optimizer.param_groups: pg["lr"] = lr
+        optimizer.zero_grad(set_to_none=True)
+
+        total_loss = 0.0
+        for _ in range(accum_steps):
+            xb, yb = get_batch(train_ids, batch_size, block_size)
+            with autocast():
+                _, loss = model(xb, yb)
+            if use_amp:
+                scaler.scale(loss / accum_steps).backward()
+            else:
+                (loss / accum_steps).backward()
+            total_loss += loss.item()
+
+        if use_amp:
+            scaler.unscale_(optimizer)
+        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+        if use_amp:
+            scaler.step(optimizer); scaler.update()
+        else:
+            optimizer.step()
+
+        step += 1; pbar.update(1)
+        if step % log_interval == 0 or step == 1:
+            pbar.set_postfix(train=f"{total_loss/accum_steps:.3f}", lr=f"{lr:.2e}")
+
+        if step % eval_every == 0:
+            vl = eval_loss()
+            best_val = min(best_val, vl)
+            print(f"     eval loss {vl:.3f} | best {best_val:.3f}")
+
+    pbar.close()
+    elapsed = time.time() - start
+    print(f"     done in {elapsed:.1f}s | best val {best_val:.3f}")
+
+    # Save model + config
+    ckpt_path = SAVE_DIR / "tinygpt.pt"
+    torch.save(model.state_dict(), ckpt_path)
+    (SAVE_DIR / "model_config.json").write_text(json.dumps({
+        "vocab_size": int(vocab_size),
+        "n_layer": n_layer, "n_head": n_head, "n_embd": n_embd,
+        "block_size": block_size, "dropout": dropout
+    }, indent=2))
+    print(f"[saved] weights → {ckpt_path}")
+    return model
+
+# --------------------------
+# Sampling helper
+# --------------------------
+def sample_chat(sp: spm.SentencePieceProcessor, model: TinyGPT, prompt: str, max_new_tokens=200):
+    prefix = f"You: {prompt}\nBot:"
+    ids = sp.encode(prefix, out_type=int)
+    x = torch.tensor([ids], dtype=torch.long, device=DEVICE)
+    with torch.no_grad():
+        y = model.generate(x, max_new_tokens=max_new_tokens, temperature=TEMP, top_k=TOP_K, top_p=TOP_P, repetition_penalty=REP_PEN)
+    return sp.decode(y[0].tolist())
+
+# --------------------------
+# Main
+# --------------------------
+def main():
+    # Build or reuse corpus/tokenizer/ids
+    corpus_path = SAVE_DIR / "corpus.txt"
+    spm_model   = SAVE_DIR / "spm_chat.model"
+    if not corpus_path.exists():
+        corpus_path = build_corpus()
+    else:
+        print(f"[cache] using {corpus_path}")
+
+    sp = spm.SentencePieceProcessor()
+    if not spm_model.exists():
+        sp = train_spm(corpus_path)
+    else:
+        sp.load(str(spm_model))
+        print(f"[cache] using {spm_model}")
+
+    enc_train = SAVE_DIR / "train_ids.pt"
+    enc_val   = SAVE_DIR / "val_ids.pt"
+    vocab_txt = SAVE_DIR / "vocab_size.txt"
+
+    if enc_train.exists() and enc_val.exists() and vocab_txt.exists():
+        train_ids = torch.load(enc_train, map_location=DEVICE)
+        val_ids   = torch.load(enc_val,   map_location=DEVICE)
+        vocab_size = int(vocab_txt.read_text())
+        print(f"[cache] loaded ids: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={vocab_size}")
+    else:
+        train_ids, val_ids, vocab_size = encode_corpus_to_ids(sp, corpus_path)
+        torch.save(train_ids, enc_train); torch.save(val_ids, enc_val)
+        vocab_txt.write_text(str(vocab_size))
+        print("[cache] saved encoded ids")
+
+    model = train_model(vocab_size, train_ids, val_ids)
+
+    print("\n[6/6] Samples:\n")
+    prompts = [
+        "Give me a spicy take on AI.",
+        "Roast my messy desk.",
+        "Explain recursion like you're annoyed.",
+        "Write a satirical headline about coffee.",
+        "Give me a shower thought about umbrellas.",
+        "Tell me a one-liner about deadlines.",
+        "Stay in Shakespeare mode and flatter me.",
+        "Reply sarcastically to: I love meetings.",
+        "What's a good way to say no to a meeting politely?",
+        "Roleplay as my productivity coach for two turns.",
+    ]
+    out_path = SAVE_DIR / "samples.txt"
+    with out_path.open("w", encoding="utf-8") as f:
+        for p in prompts:
+            txt = sample_chat(sp, model, p, max_new_tokens=200)
+            print("----\n" + txt)
+            f.write("----\n" + txt + "\n")
+    print(f"\n[saved] samples → {out_path}")
+
+if __name__ == "__main__":
+    main()