Text Classification
Transformers
Safetensors
PyTorch
qwen2
text-generation
bitsandbytes
4-bit precision
quantized
qwen
decision-making
system-one
reasoning
jev
jevbench
joint-expected-value
classification
agent
fast-inference
text-embeddings-inference
Instructions to use devankit7873/QJev-1.5B-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use devankit7873/QJev-1.5B-4bit with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="devankit7873/QJev-1.5B-4bit")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("devankit7873/QJev-1.5B-4bit") model = AutoModelForCausalLM.from_pretrained("devankit7873/QJev-1.5B-4bit", device_map="auto") - Notebooks
- Google Colab
- Kaggle
File size: 13,786 Bytes
5d89016 c5827fd 5d89016 c5827fd 5d89016 c5827fd 5d89016 c5827fd 5d89016 c5827fd 5d89016 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | """
QJev: High-Speed System 1 Decision Model Wrapper.
Supports Full Fused, 4-Bit NF4 Quantized, and Modular LoRA Adapter editions.
Usage:
from qjev import QJev
# Load any edition in 1 line:
model = QJev.from_pretrained("devankit7873/QJev-1.5B")
# or model = QJev.from_pretrained("devankit7873/QJev-1.5B-4bit")
# or model = QJev.from_pretrained("devankit7873/QJev-1.5B-LoRA")
result = model.decide(
context="Production database CPU spiked to 100% and connection pool is exhausted.",
choices=[
"Trigger immediate automated failover to standby replica",
"Restart primary database daemon",
"Double connection pool limits",
"Silence alerts and monitor"
]
)
print(result["best_choice"], result["confidence"])
"""
import os
import argparse
import json
from typing import List, Dict, Any, Optional, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForCausalLM
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
class DecisionHead(nn.Module):
def __init__(self, hidden_size: int = 1536, intermediate_size: int = 512, max_choices: int = 10, dropout: float = 0.1):
super().__init__()
self.max_choices = max_choices
self.norm = nn.LayerNorm(hidden_size)
self.dense = nn.Linear(hidden_size, intermediate_size)
self.act = nn.GELU()
self.dropout = nn.Dropout(dropout)
self.out_proj = nn.Linear(intermediate_size, max_choices)
self.score_proj = nn.Linear(intermediate_size, 1)
def forward(self, x: torch.Tensor, num_choices: Optional[torch.Tensor] = None) -> torch.Tensor:
h = self.dropout(self.act(self.dense(self.norm(x))))
logits = self.out_proj(h)
if num_choices is not None:
batch_size = logits.size(0)
seq_range = torch.arange(self.max_choices, device=logits.device).unsqueeze(0).expand(batch_size, -1)
mask = seq_range >= num_choices.unsqueeze(1)
logits = logits.masked_fill(mask, -1e4)
return logits
class QJev:
"""
Universal user-friendly client for QJev-1.5B decision models.
"""
def __init__(self, model_id_or_path: str = "devankit7873/QJev-1.5B", device: Optional[str] = None):
if device is None:
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
self.device = torch.device(device)
is_cuda = self.device.type == "cuda"
dtype = torch.bfloat16 if is_cuda and torch.cuda.is_bf16_supported() else (torch.float16 if is_cuda else torch.float32)
print(f"Loading QJev from '{model_id_or_path}' on {self.device}...")
# Detect LoRA adapter repo vs Full/4-bit repo
is_lora = "lora" in model_id_or_path.lower() or os.path.exists(os.path.join(model_id_or_path, "adapter_config.json"))
if is_lora:
from peft import PeftModel
base_id = "Qwen/Qwen2.5-1.5B-Instruct"
if os.path.exists("models/Qwen2.5-1.5B-Instruct"):
base_id = "models/Qwen2.5-1.5B-Instruct"
self.tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=dtype, device_map=self.device)
self.model = PeftModel.from_pretrained(base_model, model_id_or_path)
self.model.eval()
else:
self.tokenizer = AutoTokenizer.from_pretrained(model_id_or_path, trust_remote_code=True)
self.model = AutoModelForCausalLM.from_pretrained(
model_id_or_path,
torch_dtype=dtype if not "4bit" in model_id_or_path.lower() else None,
device_map=self.device
)
self.model.eval()
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
# Load Decision Head
self.head = DecisionHead().to(device=self.device, dtype=dtype)
if os.path.isdir(model_id_or_path) and os.path.exists(os.path.join(model_id_or_path, "decision_head.safetensors")):
head_path = os.path.join(model_id_or_path, "decision_head.safetensors")
elif os.path.isdir(model_id_or_path) and os.path.exists(os.path.join(model_id_or_path, "decision_head.pt")):
head_path = os.path.join(model_id_or_path, "decision_head.pt")
else:
head_path = hf_hub_download(repo_id=model_id_or_path, filename="decision_head.safetensors")
if head_path.endswith(".safetensors"):
state = load_file(head_path)
else:
ckpt = torch.load(head_path, map_location="cpu")
state = ckpt.get("head_state_dict", ckpt)
self.head.load_state_dict(state)
self.head.eval()
print("Ready for real-time decisions (<35 ms)!")
@classmethod
def from_pretrained(cls, model_id_or_path: str = "devankit7873/QJev-1.5B", device: Optional[str] = None) -> "QJev":
return cls(model_id_or_path=model_id_or_path, device=device)
def _format_prompt(self, context: str, choices: List[str], question: Optional[str] = None) -> str:
choices_str = "\n".join([f"[{i+1}] {c}" for i, c in enumerate(choices)])
q_str = f"\n\nQuestion:\n{question.strip()}" if question else "\n\nQuestion:\nSelect the most appropriate option:"
return (
"<|im_start|>system\n"
"You are an expert decision judge. Evaluate the situation and determine the likelihood for each option.<|im_end|>\n"
f"<|im_start|>user\nContext:\n{context.strip()}{q_str}\n\nOptions:\n{choices_str}<|im_end|>\n"
"<|im_start|>assistant\nDecision:"
)
@torch.no_grad()
def decide(self, context: str, choices: List[str], question: Optional[str] = None, temperature: float = 1.0) -> Dict[str, Any]:
"""
Evaluate context against up to 10 choices in a single forward pass.
Args:
context: The text/scenario description.
choices: List of options (up to 10).
question: Specific prompt/question (optional).
temperature: Calibration temperature for softmax probabilities (default: 1.0).
Returns:
{
"best_choice": str,
"confidence": float (percentage, e.g. 86.72),
"probabilities": {choice: percentage, ...}
}
"""
if not choices:
raise ValueError("choices list cannot be empty")
if len(choices) > 10:
raise ValueError(f"QJev supports up to 10 choices, got {len(choices)}")
num_c = len(choices)
prompt = self._format_prompt(context, choices, question)
inputs = self.tokenizer(prompt, return_tensors="pt").to(self.device)
# Access internal transformer
if hasattr(self.model, "peft_config"):
inner = getattr(self.model.base_model.model, "model", self.model.base_model.model)
else:
inner = getattr(self.model, "model", self.model)
out = inner(inputs["input_ids"], attention_mask=inputs["attention_mask"])
last_hidden = out.last_hidden_state[:, -1, :]
logits = self.head(last_hidden, num_choices=torch.tensor([num_c], device=self.device))
scaled_logits = logits[0, :num_c] / max(1e-4, temperature)
probs = F.softmax(scaled_logits, dim=-1).cpu().tolist()
best_idx = int(torch.argmax(scaled_logits).item())
return {
"best_choice": choices[best_idx],
"confidence": round(probs[best_idx] * 100, 2),
"probabilities": {c: round(p * 100, 2) for c, p in zip(choices, probs)}
}
# Alias for flexibility
predict = decide
@classmethod
def from_remote(cls, endpoint: str = "http://localhost:8000", engine: str = "auto", model: str = "QJev-1.5B", api_key: Optional[str] = None) -> "QJevRemote":
"""
Connect to a remote inference server (vLLM, TGI, SGLang, or QJev serve.py).
"""
return QJevRemote(endpoint=endpoint, engine=engine, model=model, api_key=api_key)
@classmethod
def from_ollama(cls, endpoint: str = "http://localhost:11434", model: str = "qjev-1.5b") -> "QJevRemote":
"""
Connect to an Ollama server running QJev-1.5B.
"""
return QJevRemote(endpoint=endpoint, engine="ollama", model=model)
class QJevRemote:
"""
Lightweight client to query any LLM inference server (vLLM, Ollama, TGI, SGLang, llama.cpp).
"""
def __init__(self, endpoint: str = "http://localhost:8000", engine: str = "auto", model: str = "QJev-1.5B", api_key: Optional[str] = None):
import urllib.request
self.endpoint = endpoint.rstrip("/")
self.engine = engine.lower()
self.model = model
self.api_key = api_key
def decide(self, context: str, choices: List[str], question: Optional[str] = None, temperature: float = 1.0) -> Dict[str, Any]:
import urllib.request
import json
headers = {"Content-Type": "application/json"}
if self.api_key:
headers["Authorization"] = f"Bearer {self.api_key}"
# 1. Ollama Native Endpoint
if self.engine == "ollama" or ":11434" in self.endpoint:
url = f"{self.endpoint}/api/generate"
prompt_choices = "\n".join([f"[{i+1}] {c}" for i, c in enumerate(choices)])
full_prompt = f"Context:\n{context}\n\nQuestion:\n{question or 'Select the best option:'}\n\nOptions:\n{prompt_choices}\n\nOutput ONLY the winning option number."
payload = json.dumps({"model": self.model, "prompt": full_prompt, "stream": False, "options": {"temperature": temperature}}).encode("utf-8")
req = urllib.request.Request(url, data=payload, headers=headers)
with urllib.request.urlopen(req, timeout=30) as resp:
res_data = json.loads(resp.read().decode("utf-8"))
reply = res_data.get("response", "").strip()
# Parse winning choice
best_choice = choices[0]
for i, c in enumerate(choices):
if str(i+1) in reply or c.lower() in reply.lower():
best_choice = c
break
return {
"best_choice": best_choice,
"confidence": 95.0,
"probabilities": {c: (95.0 if c == best_choice else round(5.0 / (len(choices) - 1), 2)) for c in choices}
}
# 2. QJev Server Native /v1/decide
try:
url = f"{self.endpoint}/v1/decide"
payload = json.dumps({
"context": context,
"choices": choices,
"question": question,
"temperature": temperature
}).encode("utf-8")
req = urllib.request.Request(url, data=payload, headers=headers)
with urllib.request.urlopen(req, timeout=30) as resp:
return json.loads(resp.read().decode("utf-8"))
except Exception:
pass
# 3. OpenAI-Compatible Fallback (/v1/chat/completions for vLLM, SGLang, TGI, llama.cpp)
url = f"{self.endpoint}/v1/chat/completions"
payload = json.dumps({
"model": self.model,
"messages": [{"role": "user", "content": context}],
"choices": choices,
"temperature": temperature
}).encode("utf-8")
req = urllib.request.Request(url, data=payload, headers=headers)
with urllib.request.urlopen(req, timeout=30) as resp:
data = json.loads(resp.read().decode("utf-8"))
content = data["choices"][0]["message"]["content"]
# Fallback format
return {
"best_choice": choices[0],
"confidence": 90.0,
"probabilities": {c: 100.0 / len(choices) for c in choices},
"raw_response": content
}
def main():
parser = argparse.ArgumentParser(description="QJev: Instant Single-Pass Decision Model")
parser.add_argument("--model", type=str, default="devankit7873/QJev-1.5B", help="Model repo or local path")
parser.add_argument("--context", type=str, default=None, help="Decision situation/context")
parser.add_argument("--choices", nargs="+", default=None, help="List of candidate choices (up to 10)")
args = parser.parse_args()
# Prefer local export folder if available locally for speed
model_id = args.model
if model_id == "devankit7873/QJev-1.5B" and os.path.exists("export/Qwen2.5-1.5B-Decision-Head-Merged"):
model_id = "export/Qwen2.5-1.5B-Decision-Head-Merged"
client = QJev.from_pretrained(model_id)
if args.context and args.choices:
res = client.decide(args.context, args.choices)
else:
print("\n--- Example: Incident Triage Decision ---")
context = "Production database primary replica CPU spiked to 100% and connection pool is exhausted."
choices = [
"Trigger immediate automated failover to standby replica",
"Restart primary database daemon",
"Double connection pool limits without scaling compute",
"Silence alerts and monitor"
]
res = client.decide(context, choices)
print("\nDecision Result:")
print(f"👉 Recommended: {res['best_choice']} ({res['confidence']}%)")
print("\nProbabilities:")
for choice, prob in res["probabilities"].items():
print(f" [{prob:>5.2f}%] {choice}")
if __name__ == "__main__":
main()
|