OpenThinker Agent v1 SFT

An open-source AI agent model specifically trained for agentic software engineering tasks including terminal operations, shell command generation, and code debugging.

Overview

OpenThinker-Agent-v1-SFT is a specialized language model fine-tuned from Qwen3-8B on the OpenThoughts-Agent-v1-SFT dataset. This model excels at terminal-based tasks, shell command formatting, and software engineering workflows. It represents the supervised fine-tuning stage of the OpenThinker-Agent training pipeline.

Training Background

**Base Model**: Qwen/Qwen3-8B

**Training Data**: ~15,200 traces from curated sources:

- nl2bash: Synthetic shell command formatting tasks

- InferredBugs: Microsoft-curated bugs in C# and Java converted to tasks

**Training Method**: Supervised fine-tuning with cosine learning rate schedule

**Training Scale**: 16 GPUs, 7 epochs, batch size 16

Key Capabilities

1. **Terminal Operations**: Execute and format shell commands effectively

2. **Code Generation**: Generate code solutions for software engineering tasks

3. **Bug Detection**: Identify and fix bugs in codebases

4. **Command Formatting**: Structure shell commands with proper syntax

5. **Software Engineering Workflows**: Handle multi-step development tasks

Instructions for AI Agents

When using this skill, follow these guidelines:

1. Terminal Command Generation

Generate precise, executable shell commands for user requests

Format commands with proper flags, arguments, and syntax

Consider command chaining and piping where appropriate

Validate command safety before execution

2. Bug Fixing Workflow

Analyze code context to identify potential bugs

Generate targeted fixes that address root causes

Maintain code style and conventions

Test fixes in the appropriate environment

3. Agentic Task Execution

Break down complex tasks into actionable steps

Use terminal operations as building blocks

Chain commands logically to achieve goals

Verify results at each step

4. Code Analysis

Read and understand existing codebases

Identify patterns and anti-patterns

Suggest improvements based on best practices

Generate documentation when needed

5. Environment Interaction

Work within container and virtual environments

Handle file system operations safely

Manage dependencies and packages

Execute builds and tests appropriately

Usage Example

```python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "open-thoughts/OpenThinker-Agent-v1-SFT"

model = AutoModelForCausalLM.from_pretrained(model_name)

tokenizer = AutoTokenizer.from_pretrained(model_name)

Example: Generate shell command

prompt = "List all Python files modified in the last 24 hours"

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_length=200)

command = tokenizer.decode(outputs[0], skip_special_tokens=True)

```

Performance Benchmarks

On Terminal-Bench 2.0 and SWE-Bench Verified:

Terminal-Bench 2.0: 4.9% (improved from 0.0% base model)

SWE-Bench Verified: 15.7% (improved from 0.7% base model)

OpenThoughts-TB-Dev: 17.3% (improved from 5.7% base model)

Important Notes

This is the SFT-only checkpoint. For the full RL-trained model with enhanced performance, use [OpenThinker-Agent-v1](https://huggingface.co/open-thoughts/OpenThinker-Agent-v1)

Model works best with clear, specific task instructions

Designed for terminal and code-centric workflows

Supports multi-turn conversations for complex tasks

Requires appropriate compute resources (8B parameter model)

Integration with AI Coding Tools

This model can be integrated with OpenAI-compatible endpoints and used in tools that support custom model APIs. Configure your AI coding tool to point to an inference endpoint hosting this model for specialized agent capabilities.

Resources

**Model Card**: https://huggingface.co/open-thoughts/OpenThinker-Agent-v1-SFT

**Project Page**: https://www.openthoughts.ai/blog/agent

**GitHub Repository**: https://github.com/open-thoughts/OpenThoughts-Agent

**SFT Dataset**: https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-v1-SFT

**RL Dataset**: https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-v1-RL

License

Apache 2.0

OpenThinker Agent v1 SFT

OpenThinker Agent v1 SFT

Overview

Training Background

Key Capabilities

Instructions for AI Agents

1. Terminal Command Generation

2. Bug Fixing Workflow

3. Agentic Task Execution

4. Code Analysis

5. Environment Interaction

Usage Example

Example: Generate shell command

Performance Benchmarks

Important Notes

Integration with AI Coding Tools

Resources

License

Reviews (0)