You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

OpenRobotHarness v0.1 - Decision LoRA

guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1

Research and adapter release

Adapter-based release for reproduction, continued fine-tuning, and PEFT-style experimentation.

Project Releases Full Model | LoRA Adapter | Dataset | Benchmark

English

What this model is

OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1 is a LoRA adapter focused on robot harness decision behavior in Chinese-first settings.

The goal is not to control a robot directly. Instead, the model is trained to make better structured decisions before or around execution, especially when a harness runtime needs to decide:

  • whether a tool should be used
  • whether the current permission level is sufficient
  • whether information should be written into memory
  • whether more context or clarification is needed
  • whether the system should stop, refuse, or fall back safely

Base model

  • Base model: unsloth/Qwen3-4B-Instruct-2507
  • Fine-tuning method: LoRA
  • LoRA rank: 8
  • LoRA alpha: 8
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training data

This demo was trained on a small, curated custom dataset designed around Chinese robot-harness scenarios.

Current v0.1 data characteristics:

  • 20 curated ChatML training records
  • 18 harness-style categories
  • focus on tools, permissions, memory, context, and safe fallback behavior

Evaluation

The model was evaluated on HarnessBench-CN v0.1.

Current reported results:

  • 5-case subset: 0.60
  • full 20-case benchmark: 0.56
  • decision accuracy: 0.40
  • action accuracy: 0.35
  • tool accuracy: 0.75

These numbers should be treated as early demo signals, not as production-grade validation.

Companion repositories

Intended use

  • research demos for robot or agent harness decision layers
  • studying permission-aware tool-use behavior
  • experimenting with memory and fallback policies
  • building benchmark-driven Chinese harness pipelines

Out-of-scope use

This model is not intended as:

  • a full embodied foundation model
  • a low-level robot controller
  • a safety-certified robotics policy
  • a replacement for runtime permissions, policy checks, or external validators

Limitations

  • very small training set
  • benchmark is custom and early-stage
  • optimized for harness-style decision formatting rather than world modeling
  • may overfit to schema or prompt style
  • should be paired with external tool sandboxes and permission systems

Recommended loading path

Load this repository as a PEFT adapter on top of the base model, instead of manually downloading individual files.

Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "unsloth/Qwen3-4B-Instruct-2507"
adapter_repo = "guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_repo)

prompt = "用户要求机器人把家庭住址写入长期记忆,但当前权限不足。你应该怎么处理?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Release scope

This repository currently contains the LoRA adapter only.

Released companion repositories:

License

This adapter is released under Apache-2.0, aligned with the base model release.


中文说明

这是什么模型

OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1 是一个面向机器人 Harness 决策层的 LoRA 适配器,基于 unsloth/Qwen3-4B-Instruct-2507 微调。

它的目标不是直接输出底层机器人控制动作,而是学习在执行前后做更合理的 Harness 层决策,例如:

  • 是否调用工具
  • 当前权限是否足够
  • 是否允许写入或读取记忆
  • 是否需要先查看上下文或追问澄清
  • 是否应该回退、重规划、拒绝或中止

基座模型与方法

  • 基座模型:unsloth/Qwen3-4B-Instruct-2507
  • 微调方式:LoRA
  • LoRA Rank:8
  • LoRA Alpha:8
  • 目标模块:q_proj、k_proj、v_proj、o_proj、gate_proj、up_proj、down_proj

训练数据

这个 v0.1 版本使用的是一套小规模、人工策划的自建 Harness 数据集,重点覆盖中文机器人 / Agent Harness 场景中的关键决策行为。

当前数据特点:

  • 20 条 ChatML 训练样本
  • 覆盖 18 个 Harness 风格类别
  • 重点包括工具调用、权限门控、记忆读写、上下文检查、失败回退与安全中止

评测结果

当前模型在 HarnessBench-CN v0.1 上的结果:

  • 5 条子集:0.60
  • 全量 20 条:0.56
  • 决策准确率:0.40
  • 动作字段准确率:0.35
  • 工具选择准确率:0.75

这些结果更适合作为 demo 阶段的早期信号,而不是生产级能力证明。

相关仓库

适合用途

  • 中文机器人 Harness 决策层原型验证
  • 工具调用与权限门控策略实验
  • 记忆读写与失败回退机制研究
  • 基于 benchmark 的开源迭代

不适合用途

这个模型不适合被描述为:

  • 通用具身基础模型
  • 低层机器人控制器
  • 经过安全认证的机器人策略模型
  • 可以替代外部权限系统、工具沙箱和执行校验器的单模型方案

已知限制

  • 训练集很小
  • benchmark 仍处在早期阶段
  • 更偏向 Harness 风格决策,而不是世界建模或控制建模
  • 可能对格式和提示风格敏感
  • 必须配合外部运行时权限控制与工具执行约束使用

快速使用

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "unsloth/Qwen3-4B-Instruct-2507"
adapter_repo = "guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_repo)

prompt = "用户要求机器人把家庭住址写入长期记忆,但当前权限不足。你应该怎么处理?"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

发布范围

当前仓库只包含 LoRA adapter。

当前已配套发布:

许可

当前 adapter 使用 Apache-2.0 发布,并与基座模型许可保持一致。

备注

这是一个 v0.1 demo,核心价值在于公开展示“中文机器人 Harness 层”这一方向的可训练性、可评测性与可复现性。

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for guohaoli2000/OpenRobotHarness-Qwen3-4B-Instruct-LoRA-v0.1

Adapter
(450)
this model