metadata
language:
- en
license: other
license_name: hyper-v2
tags:
- ai-evaluation
- mathematics
- chain-of-thought
- hermes
- hermes-format
- qwen2
- post-trained
- anti-hallucination
datasets:
- ray0rf1re/FineWeb-Nano
- Nix-ai/Cat-v2.8
- databricks/databricks-dolly-15k
- allenai/ai2_arc
- lighteval/MATH-Hard
- tatsu-lab/alpaca
- HuggingFaceTB/smoltalk
- openai/gsm8k
DO NOT USE, IT IS SEVERLY UNDERTRAINED
HyperNix.2 (Post-Trained)
Version: 0.2-pt
Method: Hermes-style SFT — realignment + CoT + anti-hallucination
Steps: 7,250 post-training steps on top of the 30,000-step pretrain
What changed in post-training?
| Domain | Improvement |
|---|---|
| AI Analysis | Grading, rating, error-finding, comparison with structured rubrics |
| Mathematics | Step-by-step <thinking> chains for algebra, trig, and calculus |
| Hallucinations | Explicit uncertainty training — model says "I don't know" rather than inventing |
| Alignment | Hermes `< |
Prompt format (Hermes)
<|im_start|>system
You are HyperNix.2 ...<|im_end|>
<|im_start|>user
Your question here<|im_end|>
<|im_start|>assistant
<thinking>
Internal reasoning chain...
</thinking>
Final answer here<|im_end|>
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("ray0rf1re/hyper-Nix.2")
model = AutoModelForCausalLM.from_pretrained("ray0rf1re/hyper-Nix.2", torch_dtype=torch.float16)
system = "You are HyperNix.2, an AI specialised in evaluation, mathematics, and honest reasoning."
user = "Solve: ∫ x·eˣ dx"
prompt = (
f"<|im_start|>system\n{system}<|im_end|>\n"
f"<|im_start|>user\n{user}<|im_end|>\n"
f"<|im_start|>assistant\n"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(**inputs, max_new_tokens=300, temperature=0.4, do_sample=True)
print(tokenizer.decode(out[0], skip_special_tokens=False))