This model is small. Well, that's an understatement. But welcome to the world of tiny language models.
StopAskingQuestionsMini is a six-hundred and fifty-six thousand parameter language model trained on roughly 23 million tokens of questions without answers. That may sound counterintuitive:
There is no practical reason for doing so. However, this model wasn't built for practical use, it was built to answer the ongoing question that I am trying to answer:
This project, or any of our projects, don't truly answer this - because every day, there is always a new advancement. For example, DeepSeek created
Engram, a novel architecture component that increases knowledge storage at very low compute cost.
Not much. It can generate partially coherent questions, and that's pretty much it.
StopAskingQuestionsMini uses a scaled down version of the
Qwen3 architecture.
StopAskingQuestionsMini trained on 23 million tokens of questions for two epochs with a batch size of 16.
We benchmarked our model against GPT-2, SmolLM-135M, and Qwen3-0.6B-Base on a question generation task:
Each model generated two to three hundred continuations of the prefix
Question:.
Qwen3-32B scored each one using a decimal grading system (0.0 to 1.0).
Our model generated the second highest number of coherent questions with less parameters than most character level RNNs.
Unfortunately, there is no practical use case as we stated earlier, but here are some interesting ideas:
1# =============================================================================
2# Inference
3# =============================================================================
4
5MODEL_DIR = "harley-ml/StopAskingQuestionsMini-656k" # path
6TOKENIZER_PATH = "harley-ml/StopAskingQuestionsMini-656k"
7
8# --- Generation settings ---
9PROMPT = "Question:" # prompt
10MAX_NEW_TOKENS = 96
11TEMPERATURE = 1.0
12TOP_P = 0.95
13TOP_K = 50
14REPETITION_PENALTY = 1.1
15DO_SAMPLE = True
16
17# =============================================================================
18
19import torch
20from pathlib import Path
21from transformers import (
22 AutoModelForCausalLM,
23 PreTrainedTokenizerFast,
24 AddedToken,
25)
26
27# ---------------------------------------------------------------------------
28# Device
29# ---------------------------------------------------------------------------
30
31device = (
32 "cuda" if torch.cuda.is_available() else
33 "mps" if torch.backends.mps.is_available() else
34 "cpu"
35)
36print(f"Device : {device}")
37
38# ---------------------------------------------------------------------------
39# Tokenizer (mirrors training setup)
40# ---------------------------------------------------------------------------
41
42def load_tokenizer(path: str):
43 p = Path(path).resolve()
44 if not p.exists():
45 raise FileNotFoundError(f"Tokenizer not found: {p}")
46 tok = PreTrainedTokenizerFast(tokenizer_file=str(p))
47 specials = {}
48 if tok.bos_token is None: specials["bos_token"] = AddedToken("<|bos|>", special=True)
49 if tok.eos_token is None: specials["eos_token"] = AddedToken("<|eos|>", special=True)
50 if tok.unk_token is None: specials["unk_token"] = AddedToken("<|unk|>", special=True)
51 if tok.pad_token is None:
52 if tok.eos_token is not None:
53 tok.pad_token = tok.eos_token
54 else:
55 specials["pad_token"] = AddedToken("<|pad|>", special=True)
56 if specials:
57 tok.add_special_tokens(specials)
58 tok.padding_side = "left" # left-pad for batched generation
59 return tok
60
61print("Loading tokenizer...")
62tokenizer = load_tokenizer(TOKENIZER_PATH)
63print(f" Vocab size : {tokenizer.vocab_size}")
64print(f" BOS : {tokenizer.bos_token!r}")
65print(f" EOS : {tokenizer.eos_token!r}")
66print(f" PAD : {tokenizer.pad_token!r} (id={tokenizer.pad_token_id})")
67
68# ---------------------------------------------------------------------------
69# Model
70# ---------------------------------------------------------------------------
71
72print(f"\nLoading model from {MODEL_DIR} ...")
73model = AutoModelForCausalLM.from_pretrained(
74 MODEL_DIR,
75 dtype=torch.float16 if device == "cuda" else torch.float32,
76 low_cpu_mem_usage=True,
77)
78model.eval()
79model.to(device)
80
81total_params = sum(p.numel() for p in model.parameters())
82print(f" Parameters : {total_params:,}")
83
84# ---------------------------------------------------------------------------
85# Generation helper
86# ---------------------------------------------------------------------------
87
88def generate(
89 prompt: str = PROMPT,
90 max_new_tokens: int = MAX_NEW_TOKENS,
91 temperature: float = TEMPERATURE,
92 top_p: float = TOP_P,
93 top_k: int = TOP_K,
94 repetition_penalty: float = REPETITION_PENALTY,
95 do_sample: bool = DO_SAMPLE,
96) -> str:
97
98 bos = tokenizer.bos_token or ""
99 full_prompt = bos + prompt
100
101 inputs = tokenizer(
102 full_prompt,
103 return_tensors="pt",
104 add_special_tokens=False,
105 ).to(device)
106 inputs.pop("token_type_ids", None) # Qwen3 doesn't use this
107
108 gen_kwargs = dict(
109 max_new_tokens = max_new_tokens,
110 do_sample = do_sample,
111 repetition_penalty = repetition_penalty,
112 eos_token_id = tokenizer.eos_token_id,
113 pad_token_id = tokenizer.pad_token_id,
114 )
115 if do_sample:
116 gen_kwargs["temperature"] = temperature
117 gen_kwargs["top_p"] = top_p
118 gen_kwargs["top_k"] = top_k
119
120 with torch.inference_mode():
121 output_ids = model.generate(**inputs, **gen_kwargs)
122
123 # Strip the prompt tokens so we only return what was generated
124 prompt_len = inputs["input_ids"].shape[-1]
125 new_ids = output_ids[0][prompt_len:]
126 return tokenizer.decode(new_ids, skip_special_tokens=True)
127
128
129# ---------------------------------------------------------------------------
130# Run
131# ---------------------------------------------------------------------------
132
133if __name__ == "__main__":
134 print(f"\nPrompt : {PROMPT!r}")
135 print("-" * 60)
136
137 output = generate(PROMPT)
138
139 print("Generated:")
140 print(output)
1@misc{stopaskingquestionsmini-656k,
2 title = {StopAskingQuestionsMini-656k: Questions with No Answers},
3 author = {Harley-ml},
4 year = {2026},
5 url = {https://huggingface.co/Harley-ml/StopAskingQuestionsMini-656k}
6}
