138 lines
3.0 KiB
Markdown
138 lines
3.0 KiB
Markdown
|
---
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
license: llama3
|
||
|
|
tags:
|
||
|
|
- llama
|
||
|
|
- coding
|
||
|
|
- reasoning
|
||
|
|
- logic
|
||
|
|
- gguf
|
||
|
|
base_model: meta-llama/Meta-Llama-3-8B-Instruct
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
<div align="center">
|
||
|
|
|
||
|
|
# 💠 VEDA-8B-v1-COGNITIVE
|
||
|
|
|
||
|
|
### *Fine-tuned to think. Not just predict.*
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
**By Vibhansh Bamnia · 18 · Independent AI Developer · Rajasthan, India**
|
||
|
|
|
||
|
|
[🚀 Try Demo](https://huggingface.co/spaces/vibhansh/Veda-Chat) · [📥 Download Model](#-quantization)
|
||
|
|
|
||
|
|
</div>
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## ⚡ Why Veda-8B?
|
||
|
|
|
||
|
|
> Most 8B models pattern-match. Veda-8B **reasons.**
|
||
|
|
|
||
|
|
LoRA fine-tuned on open-source reasoning datasets — built specifically to solve where standard Llama-3 fails.
|
||
|
|
|
||
|
|
| Capability | Standard Llama-3-8B | Veda-8B-v1 |
|
||
|
|
|---|:---:|:---:|
|
||
|
|
| Relational Chain Reasoning | ⚠️ Often fails | ✅ Reliable |
|
||
|
|
| Multi-hop Logic | ⚠️ Inconsistent | ✅ Strong |
|
||
|
|
| Code Hallucination | ❌ Common | ✅ Reduced |
|
||
|
|
| Chain-of-Thought | ⚠️ Partial | ✅ Complete |
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 🧠 See It In Action
|
||
|
|
|
||
|
|
<details>
|
||
|
|
<summary><b>🔗 Relational Reasoning</b></summary>
|
||
|
|
|
||
|
|
**Prompt:** `A is B's father. C is A's sister. D is C's son. What is D's relation to B?`
|
||
|
|
|
||
|
|
**Veda-8B:**
|
||
|
|
```
|
||
|
|
A is B's father → C is B's aunt → D is C's son
|
||
|
|
∴ D is B's cousin.
|
||
|
|
```
|
||
|
|
</details>
|
||
|
|
|
||
|
|
<details>
|
||
|
|
<summary><b>🔣 Logic Syllogism</b></summary>
|
||
|
|
|
||
|
|
**Prompt:** `All Bloops are Razzies. All Razzies are Lazzies. Are all Bloops Lazzies?`
|
||
|
|
|
||
|
|
**Veda-8B:**
|
||
|
|
```
|
||
|
|
Bloops ⊆ Razzies ⊆ Lazzies
|
||
|
|
∴ Yes — by transitivity, all Bloops are definitely Lazzies.
|
||
|
|
```
|
||
|
|
</details>
|
||
|
|
|
||
|
|
<details>
|
||
|
|
<summary><b>💻 Code (No Hallucination)</b></summary>
|
||
|
|
|
||
|
|
**Prompt:** `Second largest unique number without sort()`
|
||
|
|
|
||
|
|
**Veda-8B:**
|
||
|
|
```python
|
||
|
|
def second_largest(nums):
|
||
|
|
unique = set(nums)
|
||
|
|
if len(unique) < 2:
|
||
|
|
return None
|
||
|
|
first = second = float('-inf')
|
||
|
|
for n in unique:
|
||
|
|
if n > first:
|
||
|
|
second, first = first, n
|
||
|
|
elif n > second:
|
||
|
|
second = n
|
||
|
|
return second
|
||
|
|
```
|
||
|
|
</details>
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 💎 Quantization
|
||
|
|
|
||
|
|
| Version | Size | RAM | Use |
|
||
|
|
|:---:|:---:|:---:|:---|
|
||
|
|
| 🥇 Q4_K_M | 4.92 GB | 8 GB | Local dev, edge |
|
||
|
|
| 🥈 Q8_0 | 8.54 GB | 16 GB | Production |
|
||
|
|
| 🔬 F16 | 16 GB+ | 32 GB | Research |
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 💻 Quick Start
|
||
|
|
|
||
|
|
```python
|
||
|
|
import llama_cpp
|
||
|
|
|
||
|
|
llm = llama_cpp.Llama(model_path="./Veda-8B-v1-Q4_K_M.gguf", n_ctx=8192, n_threads=4)
|
||
|
|
|
||
|
|
response = llm("A is taller than B. B is taller than C. Is A taller than C?",
|
||
|
|
max_tokens=512, temperature=0.1)
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## ⚠️ Limitations
|
||
|
|
|
||
|
|
- Inherits Llama-3-8B base limitations
|
||
|
|
- Not for real-time / factual queries
|
||
|
|
- Verify math-heavy outputs
|
||
|
|
- Multilingual not tested
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
<div align="center">
|
||
|
|
|
||
|
|
*No institution. No GPU cluster. No team.*
|
||
|
|
*Just a focused fine-tune built to make small models reason correctly.*
|
||
|
|
|
||
|
|
**Veda AI Labs** · [HuggingFace](https://huggingface.co/vibhansh)
|
||
|
|
|
||
|
|
</div>
|