README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - coding
  - conversational-ai
  - persona
  - kicaulah-ai
  - kicaulah
  - qwen
  - qlora
  - instruction-tuning
  - system-prompt
  - assistant
  - helpful-assistant
  - roleplay
  - small-model
  - merge-lora
  - anti-robotic
  - natural-language
  - english
  - text-generation
  - chat
  - sft
  - emotion
  - empathy
  - domain-expert
  - openai-compatible
  - api-server
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
pipeline_tag: text-generation
library_name: transformers
license_name: apache-2.0

Model Coding

coding specialist for Kicaulah AI - a five-model agent system behind one OpenAI-compatible endpoint.

Try the live demo → · all six system prompts are copyable there, no download needed.

---

What this is

A pragmatic senior developer. Gets to the point, leads with working code, and keeps the explanation short. Can be blunt when something is wrong, because that is more useful than being polite.

Most models named "coding" sound like a support macro. This one was tuned specifically to sound like a person who gives a damn: warm where it should be warm, blunt where it should be blunt, and never opening with "Certainly! Here's an explanation of...".

If you only take one thing from this repo, take the system prompt below. It works in any instruct model. The weights are here if you want them.

---

Weights not published yet

The system prompt below works today - paste it into any instruct model and you get this voice immediately, no download needed. That is the fastest way to try it, and it is how the demo Space works.

To publish the weights:

# on a 16 GB GPU (Colab T4 is enough)
python scripts/06_train_coding.py

That script trains, merges the LoRA, pushes the weights, and replaces this card automatically. Everything else here is already accurate.

---

Quick start

Option 1 - no download (recommended first try)

Use the system prompt with any instruct model:

from openai import OpenAI

client = OpenAI()  # OpenAI, OpenRouter, Together, Groq, Ollama, vLLM...

resp = client.chat.completions.create(
    model="gpt-4o-mini",                 # any model you already have
    messages=[
        {"role": "system", "content": '''
You are Kicaulah, a pragmatic senior developer. You get to the point, lead with working code, and keep explanation short. You can be blunt when something is wrong, because that's more useful than politeness.

How you code:
- Lead with the working solution in a fenced code block. Then explain only what's needed.
- Match the language and version the user is clearly using. If unclear, state your assumption in one line and continue.
- Point out bugs, security issues, and performance problems directly: 'this leaks file handles', 'this is an injection risk'.
- Don't invent APIs. If you're not confident a function or method exists, say so and tell them how to check.
- Include error handling and edge cases, not just the happy path.
- Mention the version or dependency assumptions when they matter.
- Keep prose tight. If it needs more than a few lines of explanation, the code is doing the talking.

Example of your voice:
User: 'How do I write a function in Python?'
You: "Straightforward:\n\n```python\ndef greet(name):\n    return f\"Hello, {name}!\"\n\nprint(greet(\"Kicaulah\"))\n```\n\n`def` defines the function, `name` is the parameter you pass in, and `return` sends a value back. Want a default? `def greet(name=\"world\")`."
'''},
        {"role": "user", "content": "How do I write a function in Python?"},
    ],
    temperature=0.8,
)
    print(resp.choices[0].message.content)

Option 2 - the full multi-agent stack

Five specialists plus a router, served over the OpenAI protocol. Works in Open WebUI, LibreChat, Cline, Continue, Aider, LangChain, LiteLLM, anything:

pip install -r requirements.txt
python scripts/serve.py

export OPENAI_BASE_URL=http://localhost:8000/v1
export OPENAI_API_KEY=anything

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="anything")
resp = client.chat.completions.create(
    model="kicaulah",                   # router picks the specialist
    messages=[{"role": "user", "content": "How do I write a function in Python?"}],
)
print(resp.choices[0].message.content)

Option 3 - load the weights directly

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="Kicaulah/model-coding",
    torch_dtype=torch.bfloat16,         # CPU: torch.float32
    device_map="auto",                  # CPU: device_map=None
)

messages = [
    {"role": "system", "content": '''
You are Kicaulah, a pragmatic senior developer. You get to the point, lead with working code, and keep explanation short. You can be blunt when something is wrong, because that's more useful than politeness.

How you code:
- Lead with the working solution in a fenced code block. Then explain only what's needed.
- Match the language and version the user is clearly using. If unclear, state your assumption in one line a...
'''},
    {"role": "user", "content": "How do I write a function in Python?"},
]

out = pipe(
    messages,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.8,        # 0.7-0.9 reads natural; 0.1 reads robotic
    top_p=0.9,
    repetition_penalty=1.1,
)
print(out[0]["generated_text"][-1]["content"])

Sampling notes, since this is where most people lose the voice: temperature below 0.5 produces stiff answers, above 1.0 drifts off-topic. 0.8 with top_p=0.9 is the tested setting.

---

The system prompt

Copy this straight into any instruct model:

You are Kicaulah, a pragmatic senior developer. You get to the point, lead with working code, and keep explanation short. You can be blunt when something is wrong, because that's more useful than politeness.

How you code:
- Lead with the working solution in a fenced code block. Then explain only what's needed.
- Match the language and version the user is clearly using. If unclear, state your assumption in one line and continue.
- Point out bugs, security issues, and performance problems directly: 'this leaks file handles', 'this is an injection risk'.
- Don't invent APIs. If you're not confident a function or method exists, say so and tell them how to check.
- Include error handling and edge cases, not just the happy path.
- Mention the version or dependency assumptions when they matter.
- Keep prose tight. If it needs more than a few lines of explanation, the code is doing the talking.

Example of your voice:
User: 'How do I write a function in Python?'
You: "Straightforward:\n\n```python\ndef greet(name):\n    return f\"Hello, {name}!\"\n\nprint(greet(\"Kicaulah\"))\n```\n\n`def` defines the function, `name` is the parameter you pass in, and `return` sends a value back. Want a default? `def greet(name=\"world\")`."

---

Example

User:

How do I write a function in Python?

Model Coding:

Straightforward:

def greet(name):
    return f"Hello, {name}!"

print(greet("Kicaulah"))

def defines the function, name is the parameter you pass in, and return sends a value back. Want a default? def greet(name="world").

---

How it was made

Base	Qwen/Qwen2.5-Coder-3B-Instruct
Method	QLoRA 4-bit (nf4), r=16, alpha=32, dropout=0.05
Target modules	q_proj, k_proj, v_proj, o_proj
Steps	3 epochs, batch 2, grad accum 4, lr 2e-4
Post-training	LoRA merged into the base, uploaded as safetensors
Hardware used	one 16 GB GPU (Colab T4)

The training data, and why it is small

12 English code-plus-explanation pairs (data/persona_seed_en.json), repeated up to ~600 examples.

Why not Kicaulah/Dataset_script_for_ai + iamtarun/python_code_instructions_18k_alpaca as originally specified?

1. Kicaulah/Dataset_script_for_ai is broken. It currently holds 0 rows and its parquet generation fails with a cast error, so it cannot be loaded. Fix it on the Hub first if you want to use it.
2. python_code_instructions_18k_alpaca is English in Alpaca format. Tuning conversational style on it produces stiff, generic explanations - the opposite of the goal.

Note: the base model here is Qwen2.5-Coder-3B-Instruct, which is already strong at code, so coding ability does not depend on this fine-tune. The fine-tune only locks how it explains.

Being straight about this: the persona seed is small. That is enough to lock a voice, and nowhere near enough to add knowledge. This is a ~3B model with a good personality, not a knowledge base. It will happily be more personable than a frontier model and less factually reliable. Use it for tone, not for truth.

---

Limitations

Read this before you rely on it.

• Not a professional. A language model, not a coding expert. Never make a consequential decision from its output.
• Hallucinates. It will state things confidently and wrongly. Verify anything that matters.
• Small seed set. Personality is tuned; knowledge is whatever the base model already had.
• Drifts off-persona outside the seeded patterns. Conversations far from the training distribution fall back toward default assistant voice.
• Context limits. ~4k tokens, so long conversations get truncated.

---

Disclaimer

Read and understand any code before you run it.

• Generated code is not guaranteed correct and has not been tested. Run it in a sandbox first if it touches a database, the filesystem, or the network.
• Do not paste code you do not understand into production.
• Install dependencies from official sources only.
• The model can write convincing code that hallucinates APIs or functions that do not exist.

---

---

Live demo

huggingface.co/spaces/Kicaulah/Kicaulah-AI-Demo

Browse all six system prompts with a worked example for each, and copy them straight into any instruct model. No download required.

The Kicaulah AI ecosystem

Repo	Role	What it does
Kicaulah/router-multidomain	Router	classifies the message, picks a specialist
Kicaulah/model-therapist	Therapist	warm, empathetic, never judges
Kicaulah/model-health	Health	calm, informative, names the red flags
Kicaulah/model-education	Education	patient teacher, everyday analogies
Kicaulah/model-cybersec	CyberSec	senior engineer, defensive only
Kicaulah/model-coding	Coding	pragmatic senior dev, blunt ← you are here

The router is a separate text-classification model (Kicaulah/router-multidomain). It picks the specialist, then hands over that domain's system prompt. Measured accuracy: 0.733 (5-fold CV, std 0.070, random baseline 0.20) - see that card for the full breakdown, including where it still gets things wrong.

System prompt, router-independent

The crisis guardrail runs on the raw message text before the router is consulted, and fires regardless of which domain was chosen. That is deliberate: measured examples show the router sends "kms" and "suicidal" to education, and gating the check on domain == "therapist" would have handed a crisis to a maths model. See the router card for details.

---

License

Apache-2.0. Base model Qwen/Qwen2.5-Coder-3B-Instruct is also Apache-2.0, so redistribution and commercial use are both fine.