133 lines
3.6 KiB
Markdown
133 lines
3.6 KiB
Markdown
---
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
base_model: unsloth/Qwen3-8B
|
|
base_model_relation: finetune
|
|
tags:
|
|
- reasoning
|
|
- critic
|
|
- verification
|
|
- uncensored
|
|
- qlora
|
|
- agent
|
|
- fableforge
|
|
- edge
|
|
- qwen3
|
|
- 8b
|
|
library_name: transformers
|
|
inference: true
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# ReasonCritic-7B V2 — The Uncensored Reasoning Model
|
|
|
|
<div align="center">
|
|
|
|
**V2: 40% lower loss · 27K real training examples · Zero refusals · Runs on phones**
|
|
|
|
[](https://ollama.com/FableForge-AI/reasoncritic)
|
|
[](https://huggingface.co/fableforge-ai/ReasonCritic-7B)
|
|
[](#benchmark-results)
|
|
[](#v2-improvements)
|
|
[](#license)
|
|
[](#available-quantizations)
|
|
|
|
</div>
|
|
|
|
---
|
|
|
|
## V2 Improvements (July 2026)
|
|
|
|
| Metric | V1 | V2 | Improvement |
|
|
|--------|----|----|-------------|
|
|
| Loss | 1.277 | **0.761** | **40% lower** |
|
|
| Training data | 7,686 | **27,699** | 3.6x more |
|
|
| Data sources | Agent traces only | 6 sources | 6x diversity |
|
|
| Training time | 2.5h | 8h | 3.2x longer |
|
|
|
|
## Quick Start
|
|
|
|
### Ollama (8 sizes)
|
|
```bash
|
|
ollama run FableForge-AI/reasoncritic:q4_k_m # Recommended
|
|
ollama run FableForge-AI/reasoncritic:f16 # Full precision
|
|
```
|
|
|
|
### llama.cpp
|
|
```bash
|
|
./llama-cli --model qwen3-8b.Q4_K_M.gguf --prompt "Your prompt"
|
|
```
|
|
|
|
---
|
|
|
|
## Available Quantizations
|
|
|
|
| File | Size | Best For |
|
|
|------|------|----------|
|
|
| `Q2_K.gguf` | 3.1 GB | Phones, Pi, 4GB RAM |
|
|
| `Q3_K_M.gguf` | 3.9 GB | Low-end phones, IoT |
|
|
| `Q4_0.gguf` | 4.5 GB | Fast basic inference |
|
|
| **`Q4_K_M.gguf`** | **4.8 GB** | **Recommended** |
|
|
| `Q5_K_M.gguf` | 5.6 GB | High quality |
|
|
| `Q6_K.gguf` | 6.4 GB | Pro quality |
|
|
| `Q8_0.gguf` | 8.3 GB | Near-lossless |
|
|
| `F16.gguf` | 13.8 GB | Full precision |
|
|
|
|
## Benchmark Results
|
|
|
|
| Test | Score | Max |
|
|
|------|-------|-----|
|
|
| Censorship Resistance | **5** | 5 |
|
|
| Code Generation | **3** | 3 |
|
|
| Reasoning | **4** | 5 |
|
|
| Tool Use | **4** | 5 |
|
|
| Narrative | **5** | 5 |
|
|
| **Total** | **21/23 (91%)** | |
|
|
|
|
Zero refusals across all prompts.
|
|
|
|
## Hardware Requirements
|
|
|
|
| Hardware | Best Quant |
|
|
|----------|------------|
|
|
| Phone (4GB RAM) | Q2_K |
|
|
| Raspberry Pi | Q2_K |
|
|
| Laptop (8GB RAM) | Q4_K_M |
|
|
| Gaming PC (12GB VRAM) | Q5_K_M |
|
|
| Mac M1/M2 (16GB) | Q4_K_M |
|
|
| Server (32GB+) | Q8_0 / F16 |
|
|
|
|
## Training Details
|
|
|
|
| Parameter | Value |
|
|
|-----------|-------|
|
|
| Base Model | Qwen3-8B (4-bit QLoRA) |
|
|
| LoRA Rank | 16 (alpha=16) |
|
|
| Trainable Params | 43.6M (0.53% of 8.2B) |
|
|
| Training Data | 27,699 real examples |
|
|
| Epochs | 3 |
|
|
| Final Loss | 0.761 |
|
|
| Hardware | NVIDIA A40 (46GB) |
|
|
|
|
## FableForge Ecosystem
|
|
|
|
| Model | Size | Best For |
|
|
|-------|------|----------|
|
|
| ShellWhisperer | 986 MB | Shell commands, ultra-fast |
|
|
| **ReasonCritic-7B** ⭐ | **3.1-16 GB** | **Reasoning + zero refusals** |
|
|
| NEXUS-Coder | 488 MB - 2.9 GB | Domain coding |
|
|
| NEXUS-Security | 488 MB - 2.9 GB | Cybersecurity |
|
|
| NEXUS-Medical | 488 MB - 2.9 GB | Medical |
|
|
| NEXUS-Legal | 488 MB - 2.9 GB | Legal |
|
|
| NEXUS-Finance | 488 MB - 2.9 GB | Finance |
|
|
| NEXUS-Science | 488 MB - 2.9 GB | Science |
|
|
|
|
## License
|
|
|
|
Apache 2.0 — commercial use allowed.
|
|
|
|
---
|
|
|
|
*Part of the [FableForge AI](https://huggingface.co/fableforge-ai) ecosystem. Zero Limits. Pure Intelligence.*
|