Crawlnet/queen-v1.5

Queen v1.5 “hatchling” — Crawlnet

A small language model trained from scratch. Pretraining data and tokenizer: only pages the Crawlnet crawlers read on the crypto web (real browsers). No pretrained weights, no FineWeb, no other corpus. Chat tuning: question/answer pairs generated by deepseek-flash from those crawler pages (details below). Expect a dumb, funny, confidently-wrong model; that is the point.

Queen v1.5 is Queen v1 chat-tuned again: the same pretrained v1 weights, tokenizer and dataset; only the chat tuning is new (see What changed vs v1).

version	v1.5: version 1 of the ladder, chat-tuned again (queen-v1.5)
parameters	42.1M total, 16.9M non-embedding
architecture	nanochat GPT, depth 6, width 384, 6 heads, context 1024, vocab 16384
code	karpathy/nanochat @ 92d63d4e8bb4 (MIT) + crawlnet training/
dataset	10,521 pages exported 2026-10-04T08:42:35.913Z; after cleaning 10,518 pages, 392 domains
dataset manifest sha256	1258202cf03b6202c5c69112d5a82a34ac6d9d3cb9175fb92282ff4ebea64e32
unique training tokens	20,758,910 (+160,122 held out)
tokens trained	82,837,504 = 3.99 epochs, 4.9 tokens per non-embedding param
tokenizer	BPE, 16384 tokens, trained on the same crawler pages only
validation loss	1.5162 bits/byte on held-out crawler pages (pretraining)
chat tuning (SFT)	synthetic+openbook:deepseek-flash: 97,726 question/answer pairs generated by deepseek-flash (DeepSeek-V4.1-Flash (open weights, MIT license; DeepSeek API terms 4.2(3) allow training other models on outputs), via api.deepseek.com) from 32,676 crawler pages, plus the pairs cut from the pages. The generator was told to use only facts on each page, but its wording is in this model's chat behaviour. headings: 60,562 question/answer pairs cut out of the crawler pages themselves (FAQ questions found on pages, page headings, page titles; answers are verbatim page text). No external model was used. The only human-written words are six fixed question templates such as "Tell me about {heading}." Open book: the generated pairs were also trained once more with the crawl text they were written from in front of the question (80,471 conversations, 47% with a second page the search ranked high for the same question), so she can answer from the pages retrieved for a question.
compute	NVIDIA GeForce RTX 5080, 0.22 h wall, 0.22 GPU-hours
cost	$0.07 GPU + ~$14.83 deepseek-flash API for the chat-tuning pairs (list price)
weights	https://huggingface.co/Crawlnet/queen-v1.5

What changed vs v1

• Same pretrained v1 base model, tokenizer and pretraining dataset as Queen v1; only the chat tuning (SFT) is new.
• Chat-tuned again on the full crawl: 97,726 question/answer pairs generated from crawler pages, plus 80,471 open-book conversations (each pair again, with the crawl text it was written from in front of the question), so she answers from the pages retrieved for a question, the way the site serves her.
• blind A/B (v1 and v1.5 answering the same questions, compared without knowing which model wrote which): v1.5 better 81 times, v1 better 16 times.
• LLM-graded questions, with retrieved pages (the fixed questions of queen_train/judge_questions.json, each answered with the 2 pages retrieval finds for it, graded by an external LLM): mean correctness 0.26 -> 0.83 (of 2), about 3x; on topic 50% -> 89%.

Base: Crawlnet/queen-v1 (Queen v1, the same pretraining dataset: manifest sha256 1258202cf03b6202c5c69112d5a82a34ac6d9d3cb9175fb92282ff4ebea64e32). The version number stays 1: v1.5 is v1 chat-tuned again, not a new step of the ladder.

Contributors

Queen v1.5 was trained on pages found by these crawlers. Owners spawned theirs by burning $CRAWLNET; the wallets are the ones that made the burn, public on-chain.

crawler	owner	pages	tokens
v4	GZUQ…F59x (burn)	192	496,732
littledog	2eek…pBky (burn)	194	440,607
ghero4	FcEE…Cin4 (burn)	153	429,198
ghero5	FcEE…Cin4 (burn)	138	428,844
spiderman	5WKN…Meic (burn)	203	425,019
mrnakamoto	2eek…pBky (burn)	192	410,263
aze	BX6f…WpEf (burn)	190	401,801
spiderman	5WKN…Meic (burn)	170	380,647
near-weaver	Br2o…hHFG (burn)	149	377,785
crawlzilla	3dmq…dLBC (burn)	174	372,907
spiderman	5WKN…Meic (burn)	154	371,033
ruso	4rif…jxnZ (burn)	169	362,160
sparsity	GZUQ…F59x (burn)	188	357,951
bigdog	2eek…pBky (burn)	161	337,956
test	zero…PPUT (burn)	218	331,261
blackandwhitedogs	2eek…pBky (burn)	162	330,107
swolecrawls	F3QF…Dfrs (burn)	162	316,290
deez	CDYb…2Xxv (burn)	192	314,699
ghero7	FcEE…Cin4 (burn)	136	304,677
ghero9	FcEE…Cin4 (burn)	134	300,943
spiderman	5WKN…Meic (burn)	172	297,870
ghero2	FcEE…Cin4 (burn)	126	291,399
smokie	8kwY…JY7G (burn)	155	288,743
aze2	BX6f…WpEf (burn)	158	280,718
onepellegrino	2eek…pBky (burn)	145	273,017
spiderman	5WKN…Meic (burn)	147	267,490
ghero3	FcEE…Cin4 (burn)	123	261,159
charlotte	2eek…pBky (burn)	150	252,818
ghero6	FcEE…Cin4 (burn)	109	250,745
spiderman	H8iW…VE3N (burn)	167	250,232
warm	6TAx…SE4b (burn)	123	249,392
ch0pper	H8iW…VE3N (burn)	123	249,313
fern	2eek…pBky (burn)	153	246,142
wilbur	2eek…pBky (burn)	135	243,970
babushka	Cj9h…Cf4C (burn)	138	243,091
ghero8	FcEE…Cin4 (burn)	111	224,138
ghero1	FcEE…Cin4 (burn)	134	222,999
lmaocry	H6GW…Hmdr (burn)	110	218,530
ghero10	FcEE…Cin4 (burn)	112	210,962
ch0pper2	H8iW…VE3N (burn)	115	208,241
yancy	DPXE…XNqJ (burn)	109	207,054
dash	H2Lt…LbYp (burn)	100	192,604
jeff	81Y8…GXSt (burn)	115	188,852
manylegs	ETgB…wfpa (burn)	114	183,661
spiderman	5WKN…Meic (burn)	86	180,939
jdubs	3v2K…GTgA (burn)	62	171,264
vinland	3cSP…oYeU (burn)	61	170,127
gigachadfuru	AbMD…52T4 (burn)	96	168,825
jack	Diu8…atMh (burn)	82	168,503
redacted002	BvVr…URcY (burn)	94	165,985
weaver	F5Vd…3hVn (burn)	80	162,320
seig	4gtD…rJMz (burn)	67	159,469
pp	3JVm…Fsm4 (burn)	81	159,331
werkle-weaver	3api…NA9T (burn)	100	157,605
pp	3JVm…Fsm4 (burn)	78	151,260
redacted003	BvVr…URcY (burn)	65	147,573
weenis	36ny…e6gG (burn)	87	147,478
oreocakester	36ny…e6gG (burn)	73	145,266
morelegs	ETgB…wfpa (burn)	70	142,642
elevatd	3MRp…31K7 (burn)	70	142,065
smokiev2	8kwY…JY7G (burn)	85	141,041
peterparker	AbMD…52T4 (burn)	82	132,506
web-reaper	Krrc…fSaT (burn)	81	130,175
redacted001	BvVr…URcY (burn)	63	121,647
spiderman	FGyn…UuQt (burn)	62	118,440
turtletheory	ipMd…U4xb (burn)	53	107,294
benny	36ny…e6gG (burn)	81	94,983
cash	H2Lt…LbYp (burn)	40	85,927
grand-cou	FGcD…k1Dc (burn)	42	56,797
spy	8zQb…eogQ (burn)	24	49,975
the queen's own crawlers (6)	—	2,078	3,778,418

Pages and tokens count what made it into this version's dataset after cleaning and dedupe. Machine-readable: contributors.json.

What went in besides crawler pages

• The nanochat training code, architecture and hyperparameter rules (no weights, no data).
• Six hand-written question templates used to turn page headings into questions (see SFT above).
• Chat-tuning question/answer pairs generated by deepseek-flash from crawler pages (see SFT above).
• Nothing else. nanochat's default SFT mixture (SmolTalk, MMLU, GSM8K) and its CORE eval were switched off.

What it says

Fixed prompts, seeded sampling (temperature 0.7, top-k 50), unedited:

What is Bitcoin?

Bitcoin is a digital money market where you can be traded, transferred, or accessed by anyone. This means you can take your money by buying it, and you can trade it for the cryptocurrencies you’re earning with it.

Who created Bitcoin?

Bitcoin was created by Satoshi Nakamoto, a founder of Bitcoin. He was the first Bitcoin software developer on the Internet.

What is a blockchain?

A blockchain is a decentralized database secured by blockchain technology that powers cryptocurrencies and manages decentralized applications. It aims to reduce the number of independent blockchain networks, as well as decentralized applications like Bitcoin, Ethereum, and Solana.

Explain proof of work.

The proof of work consensus, but that’s a proof of work. Ethereum transitioned to proof of work, with Proof of work [..n]) to follow the same proof. Since Ethereum has a proof-of-work system, it has become the most secure and proof of work, and it’s important to keep the average work of it. But as mentioned, this is the most secure form of proof of work. That’s a very important requirement for Ethereum to exist.

Explain proof of stake.

Proof of stake is a mechanism that allows an operator to run validators in a specific manner based on the amount of stake that they will be able to make a unilateral decision by the operator to be able to prove that the amount of stake that they will be in the sortition pool based on the stake in the sortition pool. The stake of the network is divided into blocks, with the total stake that it will be added to the chain after it has voted for it.

What is Ethereum?

In September, Ethereum introduced the concept of decentralized digital currency (DV.) and Ether (ETH) as an alternative to traditional financial apps. It was created by Vitalik Buterin on 2014-11-15 by Buterin on 15 April 2013.

What is a smart contract?

A smart contract is a way of building an application for a decentralized application, called a smart contract. It allows users to interact with smart contracts and interact with applications via smart contracts.

What is gas on Ethereum?

Gas is the unit of measure for transaction success. It is a measure of how the network processes transactions on the Ethereum network and how it handles its computational effort.

What is Solana?

Solana is a global digital world that launched in 2015, a globally distributed network of computers. It was created to solve the problems around the world they’re designed to be censorship-resistant, secure, and decentralized.

Why is Solana fast?

Solana is fast, so it lets you build app and side apps without managing billions of users' devices. Solana is designed for builders, vaults, and products that support dApps using Solana’s infrastructure.

Use it

chat.py in this folder needs only karpathy/nanochat (MIT) at the commit above:

pip install torch tiktoken safetensors
git clone https://github.com/karpathy/nanochat && git -C nanochat checkout 92d63d4e8bb4
PYTHONPATH=nanochat python chat.py "What is staking?"

Or load it yourself: config.json has the nanochat GPTConfig; build nanochat.gpt.GPT, load model.safetensors; the tokenizer is tokenizer.tiktoken (tiktoken rank file) + the pattern and special tokens in tokenizer.json. Chat format: <|bos|><|user_start|>...<|user_end|><|assistant_start|>.

Limitations

Tiny model, narrow data. It makes things up, including prices, dates, addresses and links. It is not financial advice and must not be used as a source of facts.