---

license: apache-2.0

base_model: Qwen/Qwen2.5-Coder-7B-Instruct

pipeline_tag: text-generation

library_name: gguf

language:

• en
• zh

tags:

• gguf
• qwen2.5
• qwen
• code
• mcp
• automation
• desktop-automation
• ollama
• lm-studio
• asb

quantized_by: weahoo

---

ASB-Qwen2.5-Coder-7B-Targeted-v2-GGUF

A Q4_K_M GGUF release of a Qwen2.5-Coder-7B-Instruct derivative specialized for Automation Skill Builder (ASB) workflows.

这是面向 Automation Skill Builder（ASB） 工作流定向训练的 Qwen2.5-Coder-7B-Instruct 衍生模型，并以 Q4_K_M GGUF 格式发布。

Automation Skill Builder / visualbuild.me is a local-first AI automation platform: record once, compile the workflow into deterministic runtime logic, and run it locally without requiring AI at runtime.

What it is designed for / 训练目标

The targeted training emphasizes:

• MCP/ASB tool selection and short action planning
• parameter discovery and schema preservation
• ASB model-wiring conventions such as literal '_sb_get' / '_sb_typed_get'
• checking 'ai_sk_wire_preview' before packaging
• recovery after tool, schema, or packaging errors
• validating packaged skills with non-default parameters
• distinguishing exit-code success from verified external effects

定向能力包括：

• MCP/ASB 工具选择与短动作规划
• 参数发现、类型与 schema 保留
• ASB 模型接线规则，例如字面量 '_sb_get' / '_sb_typed_get'
• 打包前执行 'ai_sk_wire_preview'
• 工具、schema 与打包错误恢复
• 使用非默认参数验证打包产物
• 不把退出码 0 误认为业务结果必然正确

Training summary / 训练摘要

• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Training candidates: 1,980
• Validation candidates: 220
• Epochs: 2
• LoRA rank / alpha: 32 / 64
• Train loss: 0.1829
• Evaluation loss: 0.0928
• Quantization: Q4_K_M
• GGUF size: 4,683,073,472 bytes
• SHA-256: 57a2c64aa270c9bd856e6244eaa8435831b7e178634648b4e170f82377cb22d1

The GGUF was loaded and tested successfully with llama.cpp after conversion.

Benchmark:

internal 40-case mock suite — single-tool tasks ~90% pass,

schema-preservation ~90% pass, multi-step workflows ~56% pass,

error-recovery ~50% pass, safety-boundary cases ~75% pass.

Recovery and safety are the known weak points; see "Recommended operating boundary" below.

Download and run with Ollama

Recent Ollama versions can load the Hugging Face GGUF directly:

~~~bash

ollama run hf.co/weahoo/ASB-Qwen2.5-Coder-7B-Targeted-v2-GGUF:Q4_K_M

~~~

The included Modelfile requests a 49,152-token profile, but the published Q4_K_M GGUF metadata reports qwen2.context_length = 32768. In the tested Ollama setup, the metadata limit won and a ~37.3k-token ASB request was rejected. Do not treat this file as a validated 49,152-context GGUF:

~~~bash

ollama create asb-qwen25-coder-v2 -f Modelfile

ollama run asb-qwen25-coder-v2

~~~

Verify:

~~~bash

ollama show --modelfile asb-qwen25-coder-v2

~~~

LM Studio

In LM Studio, search for:

~~~text

weahoo/ASB-Qwen2.5-Coder-7B-Targeted-v2-GGUF

~~~

Select ASB-Qwen2.5-Coder-7B-Targeted-v2-Q4_K_M.gguf, then use:

~~~text

Context length: 32768 (validated GGUF metadata; 49152 was not achieved in Ollama testing)

Temperature: 0.1

OpenAI-compatible server: enabled

Default port: 1234

~~~

Independent validation warning / 独立验证警告（2026-08-28）

A Windows + Ollama + ASB lifecycle test found limitations that materially affect standard-client compatibility:

• Native OpenAI-compatible tool calling did not pass. Correct tool names and arguments were emitted as JSON inside message.content, not in the standard message.tool_calls field. Clients that route exclusively from tool_calls will not receive these calls unless they add a custom content parser. This is a basic interoperability limitation, not only a multi-step accuracy issue.
• The published Q4_K_M GGUF reports qwen2.context_length = 32768. Setting num_ctx 49152 did not override this metadata in the tested Ollama setup. ASB Start describing produced a ~37.3k-token request and was rejected before inference.
• Short code generation failed source review after two attempts. The attempts contained missing imports/CLI handling defects and an incorrect output type.
• The independent trusted-agent lifecycle was not completed by the model. Parameter analysis correctly identified 3/3 runtime inputs plus the output, and ASB wiring, packaging, and two non-default EXE runs passed only after deterministic human correction of the generated Python.

Therefore, use this release as an experimental ASB-oriented planning/analysis component behind validation gates—not as a drop-in standard tool-calling model or an unattended end-to-end trusted agent.

中文摘要：本次独立测试发现，工具调用没有进入标准 message.tool_calls 字段；GGUF 实际上下文元数据为 32768；两次短代码生成仍未通过源码审查；完整可信代理闭环未由模型独立完成。参数分析、接线、打包及 EXE 非默认参数验收在人工确定性修正代码后通过。

Public harness and full evidence: https://github.com/weahoo/asb-model-eval

Codex + ASB execution ceiling

A separate fixed 20-case EvalPlus pilot measured Codex authoring solutions and using ASB run_code as the execution surface:

• first attempt: 17/20 (85%);
• one explicit retry: 3/3 corrected, cumulative 20/20;
• ASB evaluation results returned: 20/20;
• lost or fabricated results: 0;
• representative ASB wiring, Windows packaging, and non-default runs: 3/3.

No additional correctness loss attributable to ASB was observed within this fixed pilot. This is a strong-agent platform ceiling, not a score for this Qwen model and not a full EvalPlus leaderboard result. A completely ASB-free replay was not independently executed on the same host, so the result should not be generalized as universal equivalence.

Evidence and reproducible harness:

• https://github.com/weahoo/asb-model-eval
• https://github.com/weahoo/asb-model-eval/blob/main/results/2026-08-28-codex-asb-evalplus-pilot.md

Connect to Automation Skill Builder

Ollama:

~~~text

Base URL: http://127.0.0.1:11434/v1

Model: asb-qwen25-coder-v2:latest

API key: ollama

~~~

LM Studio:

~~~text

Base URL: http://127.0.0.1:1234/v1

Model: use the model ID displayed by LM Studio

API key: lm-studio

~~~

Recommended ASB lifecycle:

1. Generate or record the workflow.
2. Analyze parameters and confirm the complete intended schema.
3. Run ai_sk_wire_preview.
4. Package only when preview returns ok: true.
5. Execute the deliverable with parameters different from recorded/default values.
6. Verify files, UI state, API responses, or other external effects—not only process exit code.

Recommended operating boundary / 推荐边界

This model is intended as a trusted planning and code-generation component inside a validated ASB workflow. It should not yet be treated as an unrestricted, unattended desktop agent.

Recommended usage:

• keep individual planning segments short—approximately five actions when possible;
• divide long MCP recordings into smaller stages;
• retain ASB schema and wiring checks;
• require non-default parameter acceptance tests;
• require human review for destructive, external-send, credential, or permission-changing actions.

本模型适合在 ASB 校验体系内作为可信规划与代码生成组件使用，暂不建议作为不受限制、无人监督的桌面代理。

A real example / 真实案例

In a controlled comparison against a cloud model (GLM-5.1) on the same ASB wiring-convention error, GLM-5.1 corrected the issue in one round;

this local model needed two rounds and still did not converge, requiring manual correction.

This is a known capability gap on multi-round self-correction, not a hypothetical caveat.

About Automation Skill Builder

Automation Skill Builder is a local-first automation layer for desktop, browser, MCP and business-system workflows. AI can help capture and parameterize an operation, while the delivered runtime logic can execute deterministically without an AI call on every run.

• Product and downloads: visualbuild.me
• Positioning: record with AI, run without AI
• Outputs: reusable scripts, MCP skills and packaged desktop deliverables
• Platforms: Windows and macOS, with workflow support varying by integration

License and attribution

This derivative follows the Apache 2.0 license of the Qwen2.5-Coder base model. Qwen and Alibaba Cloud are not affiliated with or endorsing Automation Skill Builder. Users are responsible for evaluating generated code and automation behavior for their environment.

Public evaluation harness and reviewed reports: https://github.com/weahoo/asb-model-eval

Disclaimer

Training and evaluation results do not guarantee safe or correct execution in every environment. Always validate tool arguments, paths, permissions, generated code, packaging wiring and external effects before production use.