hungrynovalabs/ nova-pup

28 Downloads Updated 4 days ago

A compact Linux-focused 4B model built for troubleshooting, Bash, system administration, and multi-agent workflows. Trained locally by Hungry Nova Labs.

4b

ollama run hungrynovalabs/nova-pup:4b

curl http://localhost:11434/api/chat \
  -d '{
    "model": "hungrynovalabs/nova-pup:4b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='hungrynovalabs/nova-pup:4b',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'hungrynovalabs/nova-pup:4b',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

nova-pup:4b

4.5GB · 256K context window · Text · 4 days ago

nova-pup:4b

4.5GB

256K

Text

Readme

Nova Pup 4B

By Matthew Salinas Hernandez — Hungry Nova Labs LLC

A 4B-parameter Linux systems specialist and multi-agent problem solver, trained end-to-end on a single RTX 5090.

Highlights

3x its base model on a sealed closed-book Linux diagnostics exam (25% → 75%)
MMLU 73.7% (vs 74.0% base) — specialization without catastrophic forgetting
Built for agent swarms: an 8-agent pack runs in 15.5 GB VRAM at 1,539 tok/s aggregate on one RTX 5090, with zero per-agent slowdown
Playful, rigorous “toon pup” persona — step-by-step reasoning with machine-checkable final answers

Training

BF16 LoRA continued pretraining on a curated Linux corpus (3,000 steps), then a solver-curriculum SFT (1,200 steps) of machine-verified puzzles: logic grids, bash-pipeline reconstruction, verification, and task decomposition.

Benchmarks (lm-evaluation-harness, vs base model)

Benchmark	Base	Nova Pup
Linux closed-book exam	25%	75%
MMLU (0-shot)	74.0%	73.7%
ARC-Challenge	54.4%	56.6%
HellaSwag	71.7%	73.9%
GSM8K (5-shot)	77.7%	71.0%

Known limitations

GSM8K math dropped 6.7 points vs base (math was not in the training mix); pair with a calculator tool for arithmetic-heavy workloads.
TruthfulQA dipped 3.7 points; calibrate confidence accordingly.
Published config uses 8k context; long-context training is planned.
Has a distinct persona; override the system prompt for a neutral assistant.

Usage

ollama run hungrynovalabs/nova-pup:4b

Training data & licensing

Nova Pup was trained using lawfully acquired technical materials and publicly available Linux documentation. Openly licensed materials retain their respective licenses. Copyrighted materials were used only for intermediate model training and are not distributed with the model.

Base model: InternScience Agents-A1-4B (Apache-2.0). Quantization: Q8_0. Also based on Qwen3.5-4B (Apache 2.0) by the Qwen Team.

# Nova Pup 4B

**By Matthew Salinas Hernandez — Hungry Nova Labs LLC**

A 4B-parameter Linux systems specialist and multi-agent problem solver, trained
end-to-end on a single RTX 5090.

## Highlights
- **3x its base model** on a sealed closed-book Linux diagnostics exam (25% → 75%)
- **MMLU 73.7%** (vs 74.0% base) — specialization without catastrophic forgetting
- **Built for agent swarms**: an 8-agent pack runs in 15.5 GB VRAM at 1,539 tok/s
  aggregate on one RTX 5090, with zero per-agent slowdown
- Playful, rigorous "toon pup" persona — step-by-step reasoning with
  machine-checkable final answers

## Training
BF16 LoRA continued pretraining on a curated Linux corpus (3,000 steps), then a
solver-curriculum SFT (1,200 steps) of machine-verified puzzles: logic grids,
bash-pipeline reconstruction, verification, and task decomposition.

## Benchmarks (lm-evaluation-harness, vs base model)
| Benchmark | Base | Nova Pup |
|---|---|---|
| Linux closed-book exam | 25% | **75%** |
| MMLU (0-shot) | 74.0% | 73.7% |
| ARC-Challenge | 54.4% | 56.6% |
| HellaSwag | 71.7% | 73.9% |
| GSM8K (5-shot) | 77.7% | 71.0% |

## Known limitations
- GSM8K math dropped 6.7 points vs base (math was not in the training mix);
  pair with a calculator tool for arithmetic-heavy workloads.
- TruthfulQA dipped 3.7 points; calibrate confidence accordingly.
- Published config uses 8k context; long-context training is planned.
- Has a distinct persona; override the system prompt for a neutral assistant.

## Usage
    ollama run hungrynovalabs/nova-pup:4b

## Training data & licensing
Nova Pup was trained using lawfully acquired technical materials and publicly
available Linux documentation. Openly licensed materials retain their
respective licenses. Copyrighted materials were used only for intermediate
model training and are not distributed with the model.

Base model: InternScience Agents-A1-4B (Apache-2.0). Quantization: Q8_0. Also based on Qwen3.5-4B (Apache 2.0) by the Qwen Team.

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)