Qwen3-0.6b_dataclaw_mallet

Qwen3 Tiny – DataClaw Coding Agent (SFT)

A lightweight instruction-tuned version of Qwen3 Tiny trained on real-world terminal and tool-using coding conversations exported with DataClaw.

This model is optimized for:

• Command-line interaction
• Tool-using agents
• Structured JSON responses
• Coding assistant behavior
• Multi-turn terminal workflows

It is small, fast, and suitable for local deployment.

---

Model Details

Model Description

This model is a supervised fine-tune (SFT) of a Qwen3 Tiny base model on structured coding-agent conversations.

Training data consists of real interactive terminal sessions including:

• User instructions
• Assistant reasoning (when present)
• Tool calls (Bash, Read, Grep, etc.)
• Structured JSON outputs
• Multi-step agent workflows

The training format preserves conversational context using a rolling window with anchored task instructions to improve task persistence.

• Developed by: Coriana
• Shared by: Coriana
• Model type: Causal decoder-only transformer
• Base model: Qwen3 Tiny (exact base version used for fine-tuning)
• Language(s): English
• License: MIT
• Fine-tuned from: Qwen3 Tiny

---

Model Sources

• Dataset: https://huggingface.co/datasets/peteromallet/dataclaw-peteromallet
• Base model: https://huggingface.co/Qwen

---

Intended Uses

Direct Use

• Local coding assistant
• CLI automation agent
• JSON-only output agents
• Tool-calling reasoning experiments
• Research into small-model agent behavior

Downstream Use

• Fine-tuning for:
  • DevOps copilots
  • Self-hosted coding assistants
  • Structured command generators
  • MUD / terminal AI agents

Out-of-Scope Use

• General knowledge QA at scale
• Legal / medical advice
• High-accuracy factual reasoning
• Large context (> trained window)
• Safety-critical systems

This is a small model trained on narrow agentic data. It is not aligned for broad real-world deployment.

---

Bias, Risks, and Limitations

• Inherits biases from base Qwen model.
• Training data is heavily skewed toward technical / terminal workflows.
• May hallucinate tool names or system states.
• Not RLHF-aligned.
• May produce unsafe shell commands.
• Limited world knowledge compared to larger models.
• Context retention depends on inference window and formatting.

Recommendations

• Run in sandboxed environments.
• Validate generated shell commands before execution.
• Constrain output schema if using for automation.
• Do not expose directly to end users without filtering.

---

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "your-username/your-model-name"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "List what directories and files are here. Just ls, no explanation needed."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=200, temperature=0.6)

print(tokenizer.decode(output[0], skip_special_tokens=True))