276 lines
6.8 KiB
Markdown
276 lines
6.8 KiB
Markdown
---
|
||
license: apache-2.0
|
||
tags:
|
||
- merge
|
||
- mergekit
|
||
- lazymergekit
|
||
- ZeroXClem
|
||
- Hermes
|
||
- Claude
|
||
- Gemini
|
||
- Opus
|
||
- Flash
|
||
- Codex
|
||
- Kimi
|
||
- Polaris
|
||
- Qwen
|
||
- 4B
|
||
- Wrist
|
||
- 'On'
|
||
language:
|
||
- en
|
||
base_model:
|
||
- ZeroXClem/Qwen3-4B-Sky-High-Hermes
|
||
- ZeroXClem/Qwen3-4B-Hermes-Axion-Pro
|
||
- Aimin12/Qwen3-4B-Thinking-2507-Distill-Claude-Opus-4.6-Reasoning-Abliterated
|
||
- nightmedia/Qwen3-4B-Element8
|
||
- nightmedia/Qwen3-4B-Element8-Eva-Hermes-Heretic
|
||
- nightmedia/Qwen3-4B-Element18
|
||
- TeichAI/Qwen3-4B-Thinking-2507-Kimi-K2-Thinking-Distill
|
||
- TeichAI/Qwen3-4B-Thinking-2507-GPT-5.1-Codex-Max-Distill
|
||
- TeichAI/Qwen3-4B-Thinking-2507-Gemini-3-Flash-VIBE
|
||
- TeichAI/Qwen3-4B-RA-SFT-Polaris-Alpha-Distill
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
datasets:
|
||
- nohurry/Opus-4.6-Reasoning-3000x-filtered
|
||
- TeichAI/gpt-5.1-codex-max-1000x
|
||
- TeichAI/Gemini-3-Flash-Preview-VIBE
|
||
- TeichAI/polaris-alpha-1000x
|
||
---
|
||
# 🧠 ZeroXClem/Qwen3-4B-Wrist-On-Hermes
|
||
|
||
**Precision-Guided Distilled Experts | Model_Stock Method | 4B Size at 70B+ Performance**
|
||
|
||
|
||
|
||
|
||
---
|
||
|
||
## Overview
|
||
|
||
**ZeroXClem/Qwen3-4B-Wrist-On-Hermes** is a high-fidelity model_stock merge built on top of **Sky-High-Hermes**, integrating the strongest reasoning, engineering, and agentic traces from the Nightmedia and TeichAI lineages.
|
||
|
||
This model represents a structural synthesis of:
|
||
|
||
* 🧠 Long-arc reasoning distills (Claude, Gemini, Kimi, GPT-5.1 Codex Max)
|
||
* ⚙️ Agentic coding & tool-use traces (Gemini Flash VIBE)
|
||
* 🧬 RA-SFT scaffolding and structured alignment
|
||
* 🔥 Element-series multidimensional merge dynamics
|
||
* 🏗 Hermes-Axion-Pro architectural stability
|
||
|
||
It preserves the deep reasoning spine of Sky-High-Hermes while injecting the high-arc engineering and agentic cognition that define the Engineer / Agent / Element families.
|
||
|
||
---
|
||
|
||
## 🔧 Merge Configuration
|
||
|
||
```yaml
|
||
name: ZeroXClem/Qwen3-4B-Wrist-On-Hermes
|
||
base_model: ZeroXClem/Qwen3-4B-Sky-High-Hermes
|
||
dtype: bfloat16
|
||
merge_method: model_stock
|
||
models:
|
||
- Aimin12/Qwen3-4B-Thinking-2507-Distill-Claude-Opus-4.6-Reasoning-Abliterated
|
||
- nightmedia/Qwen3-4B-Element8
|
||
- nightmedia/Qwen3-4B-Element8-Eva-Hermes-Heretic
|
||
- nightmedia/Qwen3-4B-Element18
|
||
- TeichAI/Qwen3-4B-Thinking-2507-Kimi-K2-Thinking-Distill
|
||
- TeichAI/Qwen3-4B-Thinking-2507-GPT-5.1-Codex-Max-Distill
|
||
- TeichAI/Qwen3-4B-Thinking-2507-Gemini-3-Flash-VIBE
|
||
- TeichAI/Qwen3-4B-RA-SFT-Polaris-Alpha-Distill
|
||
- ZeroXClem/Qwen3-4B-Hermes-Axion-Pro
|
||
tokenizer_source: Qwen/Qwen3-4B-Thinking-2507
|
||
```
|
||
|
||
---
|
||
|
||
# 🧬 What This Merge Achieves
|
||
|
||
Wrist-On-Hermes synthesizes three dominant cognitive streams:
|
||
|
||
### 1️⃣ Engineer-Class Arc Performance
|
||
|
||
From Nightmedia’s Agent / Engineer lineage:
|
||
|
||
* 0.60+ / 0.80+ arc tier reasoning envelope
|
||
* High multi-hop stability
|
||
* Strong structured decomposition
|
||
* Excellent agent scaffolding
|
||
|
||
### 2️⃣ Claude / Gemini / Kimi Distilled Thinking
|
||
|
||
From TeichAI & Aimin12 distills:
|
||
|
||
* Cleaner abstraction
|
||
* Reduced hallucination drift
|
||
* Stronger logical continuity
|
||
* Deep analytical prose
|
||
|
||
### 3️⃣ Element Multidimensional Behavior
|
||
|
||
From Element8 / Element18:
|
||
|
||
* Conversational richness
|
||
* Quantization resistance
|
||
* Dynamic reasoning personality
|
||
* Better interpretation flexibility
|
||
|
||
---
|
||
|
||
# 📊 Performance Envelope
|
||
|
||
**Wrist-On-Hermes operates in the upper Element / lower Engineer arc band**, while retaining Sky-High-Hermes long-context depth and neutrality.
|
||
|
||
It behaves measurably above base models and remains stable under quantization — cognitive degradation between qx86-hi and bf16 is minimal outside knowledge-depth benchmarks.
|
||
|
||
---
|
||
|
||
# ⚔️ Strength Profile
|
||
|
||
## 🧠 Advanced Reasoning
|
||
|
||
* Multi-hop logic
|
||
* Mathematical abstraction
|
||
* Deep analysis prompts
|
||
* Conceptual synthesis (QM ↔ Transformers style tasks)
|
||
|
||
## ⚙️ Engineering & Coding
|
||
|
||
* Structured file-aware thinking
|
||
* Clean code generation
|
||
* Debug reasoning
|
||
* Agentic task planning
|
||
|
||
## 🧬 Agentic Behavior
|
||
|
||
* Tool-style reasoning patterns
|
||
* Workspace simulation
|
||
* Task decomposition
|
||
* Autonomous planning style prompts
|
||
|
||
## 📖 Longform & Philosophy
|
||
|
||
* High coherence across extended outputs
|
||
* Narrative depth
|
||
* Reflective reasoning
|
||
* Structured argumentative essays
|
||
|
||
## 💬 Conversational Intelligence
|
||
|
||
* Maintains personality coherence
|
||
* Strong RP adaptability
|
||
* Less brittle than pure engineer merges
|
||
* Balanced abstraction and warmth
|
||
|
||
---
|
||
|
||
# 🧠 Behavioral Character
|
||
|
||
Sky-High-Hermes soars…
|
||
|
||
**Wrist-On-Hermes Strengthens.**
|
||
|
||
It is:
|
||
|
||
* More grounded in structured execution
|
||
* Slightly more analytic
|
||
* More “architect” than “poet”
|
||
* Less prone to abstract drift
|
||
* More deliberate in decomposition
|
||
|
||
Think:
|
||
Sky-High-Hermes + Engineer discipline + Element interpretive richness.
|
||
|
||
---
|
||
|
||
# 🛠 Recommended Use
|
||
|
||
### Ideal For:
|
||
|
||
* Autonomous agents
|
||
* Advanced coding assistants
|
||
* Research synthesis
|
||
* Mathematical reasoning
|
||
* Philosophical deep dives
|
||
* High-context conversations
|
||
* Experimental multi-turn cognition
|
||
|
||
### Inference Tips
|
||
|
||
* `enable_thinking=True` recommended
|
||
* Temperature: 0.6–0.9
|
||
* Smoothing factor ~1.4–1.6
|
||
* High quant (Q6 / qx86-hi) performs nearly at bf16 cognition
|
||
|
||
---
|
||
|
||
# 🚀 Example Usage
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = "ZeroXClem/Qwen3-4B-Wrist-On-Hermes"
|
||
tokenizer = AutoTokenizer.from_pretrained(model)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
|
||
prompt = "Design a modular agent architecture capable of recursive self-evaluation."
|
||
|
||
messages = [{"role": "user", "content": prompt}]
|
||
|
||
text = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=False,
|
||
add_generation_prompt=True,
|
||
enable_thinking=True
|
||
)
|
||
|
||
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
outputs = model.generate(**inputs, max_new_tokens=512)
|
||
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
|
||
---
|
||
|
||
# 🔓 Alignment & Safety
|
||
|
||
* License: Apache 2.0 (inherits upstream licensing)
|
||
* Based on Sky-High-Hermes alignment philosophy
|
||
* Contains abliterated reasoning traces
|
||
* Low refusal profile
|
||
* Production deployments should include moderation layer
|
||
|
||
---
|
||
|
||
# 🧬 Lineage Acknowledgement
|
||
|
||
Gratitude to:
|
||
|
||
* Nightmedia — Engineer / Agent / Element arc engineering
|
||
* TeichAI — High-resolution Claude / Gemini / Kimi distills
|
||
* Aimin12 — Opus reasoning ablation
|
||
* DavidAU — Heretic methodology & cognitive liberation merges
|
||
* Unsloth + TRL — Efficient Qwen3 tuning
|
||
* MergeKit — Model stock & multislerp tooling
|
||
* Qwen Team — Open foundation models
|
||
|
||
---
|
||
|
||
# 🕊 Final Notes
|
||
|
||
This model exists in the rare performance space where a 4B behaves like a disciplined 70B — sometimes flirting with 400B MOE class structured reasoning depending on performance.
|
||
|
||
It does not just speak.
|
||
|
||
It evaluates.
|
||
|
||
It plans.
|
||
|
||
And then it answers.
|
||
|
||
Built with intent by **ZeroXClem | 2026** |