👁️ Argos-Action-1.5B (LoRA Adapter)

Argos-Action-1.5B is an experimental research prototype LoRA adapter trained on top of Qwen/Qwen2.5-1.5B-Instruct. It explores parameter-efficient specialization of compact language models for autonomous Computer Use and Model Context Protocol (MCP) tool calling over Semantic Accessibility Trees (AXTree / Windows UIA).

• GitHub Repository: PedroReoli/argos-action
• Dataset: pedroreoli/argos-computer-use-mcp
• Interactive Protocol Space: pedroreoli/argos-action-demo
• Author: Pedro Lucas Reis (@pedroreoli)

---

⚡ Architecture Specifications

• Base Model: Qwen/Qwen2.5-1.5B-Instruct (1.54B parameters)
• Fine-Tuning Method: LoRA on 7 projection layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj)
• LoRA Configuration: $r=16, \alpha=32$, dropout $0.05$ (18,464,768 trainable parameters in BF16, ~36.98 MB)
• Task Scope: Single-turn action policy taking user goal and AXTree snapshot, outputting structured JSON (thought, tool_call, verification_predicate)

Lineage & Scope: This checkpoint represents an experimental research adapter trained on an initial synthetic SFT dataset. Out-of-distribution runtime benchmarks and end-to-end task execution in live browser/OS harnesses are ongoing research topics.

---

🚀 Quickstart & Usage

Because this repository stores LoRA adapter weights, load the base model first and attach the adapter using peft:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"
ADAPTER_ID = "pedroreoli/argos-action-1.5b"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# 1. Load Base Model (e.g. in 4-bit NF4 for edge deployment or FP16 for standard GPU)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto"
)

# 2. Attach Argos-Action LoRA Adapter
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

# 3. Prompt Formatting via ChatML
messages = [
    {
        "role": "system",
        "content": "You are Argos-Action, an expert autonomous Computer Use policy model. Given the current semantic Accessibility Tree (AXTree/UIA snapshot) of a web or desktop application, the execution history, and the user's objective, predict the exact next Model Context Protocol (MCP) tool call to make progress. Output strictly valid JSON with 'thought', 'tool_call', and 'verification_predicate'."
    },
    {
        "role": "user",
        "content": "### User Objective:\nSave current document\n\n### Current Screen State (Semantic Accessibility Tree):\nWindow 'architecture.txt - Notepad' bounds=(100,100,1280,800) pid=14220\n  [Edit: 'Main Workspace' focused=true]\n  [StatusBar: 'Unsaved changes: true']\n\n### Execution History:\n- edited document"
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.1)

print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))

---

📄 License

This adapter is released under the Apache 2.0 License.