pedroreoli/argos-computer-use-mcp

👁️ Argos: Computer Use & Semantic Tool Calling (MCP) Benchmark

Argos-Action is a synthetic single-turn decision dataset designed for research into Supervised Fine-Tuning (SFT) of Small Language Models (SLMs, e.g., Qwen 2.5 1.5B, Llama 3.2) into specialized Edge Policy Agents for Autonomous Computer Use.

---

🔬 Motivation & Research Scope

Frontier general-purpose LLMs (Claude 3.5 Sonnet, GPT-4o) provide strong reasoning but face operational challenges in continuous GUI execution loops:

1. High Latency Overhead: Multi-second roundtrips per interaction slow down interactive desktop/web control.
2. Context Inflation in Real-World DOMs: In live web pages, raw HTML DOM trees often reach 40,000 to 80,000 tokens per screen.
3. Structured Tool Parsing: Unconstrained chat models can occasionally emit markdown clutter instead of valid JSON tool calls.

Theoretical vs. Synthetic Dataset Representation:

• Theoretical Contrast (Real-World DOM vs. AXTree): In live scrapers (e.g. Achilles CDP), pruning raw DOM (45,000 tokens) into accessibility trees (3,500 tokens) yields an arithmetic reduction of ~92.2% ($1 - 3500/45000$).
• This Synthetic Dataset: Contains compact, pre-structured accessibility tree fragments designed for lightweight syntax and action policy fine-tuning (averaging 87.2 tokens of AXTree per sample).

---

🏛️ Dataset Taxonomy & Tool Distribution

The dataset contains 2,000 single-turn decision samples partitioned into 1,700 training samples and 300 strictly held-out evaluation samples across five domains:

Domain	Scope & Scenarios	Train Actions	Test Actions
E-Commerce & Retail	Search bars, price filter facets, cart manipulation, scroll discovery	click, type_text	scroll (60)
Enterprise SaaS & CRM	Data grids, pipeline kanban transitions, role selection menus	click, type_text	select_option (60)
DevOps & Cloud Consoles	Cluster pod restart, pull request review & squash merge, S3 bucket ACL	click, type_text	click (60)
Native Windows UIA	File Explorer picker dialogs, print dialog duplex & copy settings	click, press_hotkey	type_text (60)
Safety & Human Escalation	2FA/SMS verification, Cloudflare Turnstile anti-bot, database wipe guard	escalate_human	escalate_human (60)

Tool Distribution Note: In this dataset version (ae274fe), scroll (60 samples) and select_option (60 samples) are present exclusively in the held-out test split to test generalization on out-of-distribution action schemas. In the training split (1,700 samples), the distribution is: click (1,035), escalate_human (340), type_text (165), and press_hotkey (160).

---

📊 Measured Token Census (Full Population of 2,000 Rows)

Census measured directly across all 2,000 rows using the Qwen/Qwen2.5-1.5B-Instruct tokenizer (tokenizers 0.23.3, add_special_tokens=False):

Span / Field	n	Min	p50	Mean	p95	Max
input Prompt	2,000	90	127	128.4	160	164
AXTree Fragment	2,000	59	83	87.2	114	116
Assistant Output JSON	2,000	71	79	85.0	112	112
Complete ChatML Sequence	2,000	262	304	307.4	366	370

• Training sequence token total (1,700 rows $\times$ 3 epochs): 1,567,626 tokens.

SHA-256 Checksums (metadata/hashes.json)

• train.jsonl (4,669,033 bytes, 1,700 rows): 382f0f953fcf7aad758147a2009b6b615d0a7fd2881cbc278828fa28d9a6476a
• test.jsonl (836,940 bytes, 300 rows): 45f2cf222247abb2536ea2f848c5f59fb0b9c94bf6d7115871c2202247ba50fc

Holdout Independence:

• Exact overlap between the 300 test inputs/goals and the training set: 0/300.
• The test set consists of 5 disjoint templates (each instantiated with variations 60 times) that never appear in the training partition.

---

📋 Data Format & Structure

Records include dual formats: Alpaca Format (instruction, input, output) and OpenAI/ChatML Format (messages):

{
  "id": "argos_train_00042",
  "domain": "enterprise_saas",
  "instruction": "You are Argos-Action, an expert autonomous Computer Use policy model...",
  "input": "### User Objective:\nUpdate opportunity status for 'Acme Corp' to 'Closed Won'\n\n### Current Screen State (Semantic Accessibility Tree):\nWindow \"Enterprise CRM - Pipeline\" bounds=(0,0,1920,1080)\n  [Card: \"Acme Corp Deal - $120,000\" id=\"card_deal_acme\"]\n  [Tab: \"Negotiation\" state=\"active\"]\n  [Button: \"Closed Won\" id=\"btn_stage_closed_won\" actions=[click]]\n  [Button: \"Closed Lost\" id=\"btn_stage_closed_lost\"]\n\n### Execution History:\n- opened deal detail for client Acme Corp",
  "output": "{\n  \"thought\": \"The user requested marking Acme Corp's deal as 'Closed Won'. The stage button is visible in the pipeline bar.\",\n  \"tool_call\": {\n    \"tool\": \"click\",\n    \"arguments\": {\n      \"element_id\": \"btn_stage_closed_won\",\n      \"button\": \"left\"\n    }\n  },\n  \"verification_predicate\": {\n    \"expected_element\": \"modal_confirm_won\",\n    \"timeout_ms\": 2000\n  }\n}",
  "messages": [
    {"role": "system", "content": "You are Argos-Action, an expert autonomous Computer Use policy model..."},
    {"role": "user", "content": "### User Objective:\n..."},
    {"role": "assistant", "content": "{\"thought\": ...}"}
  ]
}

---

🛡️ Synthetic Fixtures Disclosure

Security templates include explicit synthetic placeholders:

• Synthetic Cloudflare Ray ID: sim-cf-ray-8b29f4c3a012-synthetic (165 occurrences).
• Masked synthetic phone number: (**) *****-4040 (350 occurrences). No real user data, production IP addresses, credentials, or private tokens are present.