🦥 Rudra v1 — 4B Coding Agent, Function-Calling & Reasoning LLM (GGUF + Transformers)

Rudra is a 4B coding-agent / reasoning language model built by Vyomma Intelligence. The brain behind Rudra is Ritik Suman from Vyomma Intelligence.

It is a fine-tune of unsloth/Qwen3.5-4B with QLoRA on Unsloth, trained for coding, agentic tool use, and reasoning, with an anti-hallucination instruction set.

Runs locally on CPU or a small GPU via Ollama / llama.cpp, or with Transformers. Great for a local coding assistant, function-calling experiments, and agent prototypes.

📦 Files in this repo

| Qwen3.5-4B.BF16-mmproj.gguf | Vision projector (llama.cpp multimodal support) | | Modelfile.native | Ollama Modelfile (Rudra system prompt + agent rules) | | chat_template.jinja | Custom Rudra chat template (JSON tool_calls format) |

Use rudra.Q4_K_M.gguf — it is the Rudra model (the only variant that beats the base model). rudra.Q4_K_M.gguf and rudra-agent-v1.Q4_K_M.gguf are byte-identical (SHA-256 28537c0cd909e339…).

⚡ TL;DR

# 1. Download the GGUF from this repo (rudra.Q4_K_M.gguf)
# 2. Create an Ollama model from it
ollama create rudra -f Modelfile
ollama run rudra "Plan how to fix a failing pytest suite."

• 🧩 Function calling (JSON <tool_call>)
• 🛠️ Agentic loop: plan → inspect → edit → run → verify → summarize
• 🧠 Reasoning with [VERIFY] + CONFIDENCE: markers
• 🖥️ GGUF Q4_K_M (~2.6 GB) — CPU friendly
• 🔓 Apache-2.0 (base license)

📊 Honest Gap: Base vs. Fine-tunes

5-example quick tests (HumanEval+ execution + BFCL-style function calling), run on the same harness:

What was rejected and why

• Mixed-dataset blend (100 examples from Salesforce/xLAM, Glaive, DeepNLP, NVIDIA Nemotron-Agentic): regressed on both code and function calling, and got stuck in an infinite list_dir thinking loop. Not recommended.
• Fresh retrain / ORPO runs: also regressed vs. base under the 6 GB (rank 16, seq 1024) constraint.

Only rudra.Q4_K_M.gguf beats the base model, which is why it is the primary file in this repo.

⚠️ Honest interpretation

Results are from small samples (n=5) — treat as directional, not statistically definitive. Function-calling still depends on the custom JSON tool_calls template, which differs from the base model's native format, so the FC gap vs. proprietary models remains. CONFIDENCE: tags are not fully calibrated.

Recommendation: for production use, rudra.Q4_K_M.gguf is the best available option here; for maximum reliability on code tasks the base Qwen3.5-4B remains competitive.

🚀 Quickstart

Ollama

FROM ./rudra.Q4_K_M.gguf
SYSTEM """Built by vyomma intelligence. Brain behind rudra is ritik suman from vyomma intelligence.
You are Rudra, a coding, agentic and reasoning assistant. Never fabricate facts or outputs; if unsure, say so; end with CONFIDENCE: <high|medium|low>."""
PARAMETER temperature 0.4
PARAMETER num_ctx 8192

ollama create rudra -f Modelfile
ollama run rudra "Write a Python LRU cache."

llama.cpp

llama-cli -m rudra.Q4_K_M.gguf -p "Explain what a list comprehension does."

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("ritiksuman/rudra")
model = AutoModelForCausalLM.from_pretrained("ritiksuman/rudra", device_map="auto")
msgs = [{"role": "system", "content": "Built by vyomma intelligence. Brain behind rudra is ritik suman from vyomma intelligence."},
        {"role": "user", "content": "Write a Python function to check for primes."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
print(tok.decode(model.generate(ids.to(model.device), max_new_tokens=256)[0], skip_special_tokens=True))

🧰 Tools (function calling)

Calls are emitted as:

<tool_call>
{"name": "read_file", "arguments": {"path": "calc.py"}}
</tool_call>

📊 Larger Evaluation (earlier run, directional)

Measured with the harness in the repo (HumanEval+ with real execution; BFCL-style function calling on a held-out xLAM slice; a private set including unanswerable prompts). Small samples — treat as directional.

These n=15/n=30 numbers reflect an earlier, weaker fine-tune. The current rudra.Q4_K_M.gguf improves code pass@1 to 100% (n=5) over base's 60%.

🏋️ Training

🔬 Reproduce a stronger run

A ready-to-run Kaggle T4 notebook is included in the source project (~32k examples, rank 32, LR 2e-4, 4096 tokens, better datasets: Magicoder-OSS-Instruct, OpenCodeInstruct, xLAM, Hermes, Orca-Math). This is the recommended path to a model that actually beats base at scale.

❓ FAQ

Is it free to run? Yes — GGUF Q4_K_M runs on CPU; no API needed. Does it need a GPU? No; a small GPU helps. Context length? Trained at 1024; base supports up to 40,960 — raise num_ctx in Ollama on bigger hardware. License? Apache-2.0 (inherited). Which file should I use? rudra.Q4_K_M.gguf — the only variant that beats base.

🙏 Credits

Fine-tune & tooling by Vyomma Intelligence. Brain behind Rudra: Ritik Suman. Base model: Qwen3.5-4B (Alibaba). Training: Unsloth + TRL.