AgentForge-MultiTurn-ToolCall-5k

A commercial-grade, synthetic, multi-turn agentic tool-calling dataset for supervised fine-tuning (SFT) of LLMs on agent trajectories. 5,000 conversations, 18,481 tool calls, 30.5 % include genuine error-recovery branches — the capability most under-represented in existing open datasets.

⚠️ ACCESS & LICENSING — READ BEFORE REQUESTING

This dataset is gated. Access is granted case-by-case.

Use case	Access	What to do
Personal / academic / non-commercial research	Granted automatically on request	Click "Request access" above. Briefly describe your research.
Commercial use (training models you sell, embed in a paid product, use internally at a company with > $1M ARR or > 10 employees)	Requires a commercial license	Email contact.tahirrasool@gmail.com with subject line AgentForge Commercial License. Include: company name, intended use, deployment scale.

The Apache-2.0 license applies to non-commercial use only. Commercial use without a signed license agreement is prohibited. If you are unsure whether your use is commercial, assume it is and email contact.tahirrasool@gmail.com.

Quoted excerpts for benchmarking or academic papers are fine without a license.

Why this dataset exists

Most open tool-calling corpora (xLAM, Gorilla, ToolBench, Hermes-Function-Calling) are dominated by single-turn, success-only traces. Real agents fail. They get HTTP 500s, schema mismatches, sold-out inventory, conflicting calendar invites, expired coupons. A model fine-tuned only on happy-path traces will hallucinate recoveries instead of executing them.

AgentForge closes that gap:

Property	AgentForge	Typical open alternatives
Multi-turn trajectories	100 %	20–40 %
Error-recovery traces	30.5 %	< 5 %
Tool schemas included in every example	Yes (OpenAI function-calling format)	Sometimes
Reasoning before every tool call	Yes	Inconsistent
Domain diversity	8 domains	1–3 domains
License	Apache-2.0 (non-commercial access; commercial license on request)	Mixed, often restrictive

Dataset structure

from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/agentforge-multiturn-toolcall", token="hf_...")

Each of the 5,000 records has the following fields:

field	type	description
id	string	Unique example id, e.g. af_00001.
domain	string	One of finance, travel, ecommerce, devops, crm, calendar, email, database.
language	string	Always en in this release.
difficulty	string	easy, medium, or hard. hard ⇔ the trace includes a recovery branch.
includes_recovery	bool	Whether the trajectory includes a tool failure that the assistant recovers from.
num_turns	int	Total messages in the conversation (system + user + assistant + tool).
num_tool_calls	int	Total tool invocations in the conversation.
tools	list[dict]	OpenAI-compatible function schemas available to the assistant. (Stored as a JSON string in parquet; use json.loads to deserialize.)
conversations	list[dict]	ShareGPT-style messages: role ∈ {system, user, assistant, tool}; assistant messages may carry a tool_calls array.

Sample conversation (abridged)

{
  "id": "af_00001",
  "domain": "devops",
  "difficulty": "hard",
  "includes_recovery": true,
  "num_turns": 14,
  "num_tool_calls": 4,
  "tools": [ /* OpenAI function schemas: list_deployments, get_deployment_logs, rollback_deployment, trigger_ci_pipeline, scale_service */ ],
  "conversations": [
    {"role": "system",    "content": "You are an SRE assistant..."},
    {"role": "user",      "content": "Our auth-service in production is throwing 500s..."},
    {"role": "assistant", "content": "Listing recent deployments...",
                          "tool_calls": [{"name": "list_deployments",
                                          "arguments": {"service": "auth-service", "environment": "production"}}]},
    {"role": "tool",      "name": "list_deployments",
                          "content": "{\"deployments\": [{\"deployment_id\": \"DEP-...\", \"status\": \"failed\"}]}"},
    {"role": "assistant", "content": "Pulling error logs.",
                          "tool_calls": [{"name": "get_deployment_logs", "arguments": {...}}]},
    {"role": "tool",      "name": "get_deployment_logs",
                          "content": "{\"logs\": [\"[ERROR] Cannot connect to redis...\", \"[ERROR] OOMKilled\"]}"},
    {"role": "assistant", "content": "Attempting rollback to v2.3.1.",
                          "tool_calls": [{"name": "rollback_deployment", "arguments": {...}}]},
    {"role": "tool",      "name": "rollback_deployment",
                          "content": "{\"status\": \"failed\", \"reason\": \"artifact missing\"}"},
    {"role": "assistant", "content": "Rollback failed — re-triggering CI to rebuild.",
                          "tool_calls": [{"name": "trigger_ci_pipeline", "arguments": {...}}]},
    {"role": "tool",      "name": "trigger_ci_pipeline",
                          "content": "{\"status\": \"triggered\", \"pipeline_id\": \"CI-...\"}"},
    {"role": "assistant", "content": "Summary: failure due to OOM + Redis; rollback failed; CI rebuild running..."}
  ]
}

Coverage

Domain	Conversations	Tools
finance	625	5
travel	625	5
ecommerce	625	6
devops	625	5
crm	625	5
calendar	625	5
email	625	5
database	625	5
Total	5,000	41 unique

Aggregate statistics

• Total tool calls: 18,481
• Average turns per conversation: 10.4
• Average tool calls per conversation: 3.7
• With error-recovery branch: 1,523 (30.5 %)
• Without (happy path): 3,477 (69.5 %)
• Difficulty: easy 1,731 · medium 1,746 · hard 1,523

Intended use

1. Supervised fine-tuning (SFT) of small-to-mid LLMs (1B–14B parameters) to learn:
   • when to call a tool vs. answer from parametric knowledge,
   • how to format OpenAI-style tool calls,
   • how to interpret tool responses,
   • how to recover from tool failures (the headline differentiator),
   • how to chain multiple tools across turns to reach a goal.
2. Evaluation of agentic capability — slice the dataset by difficulty, domain, or includes_recovery.
3. Curriculum learning — start with easy (no recovery, single domain), progressively mix in medium and hard.

Out of scope

• This dataset is synthetic. Tool responses are simulated, not real API calls.
• It does not contain PII, real customer data, or copyrighted material.
• It is not a preference dataset. For DPO/IPO, use it to generate preference pairs via rejection sampling.

Provenance & generation

• Generation method: deterministic Python generator with fixed random seed (20260629).
• Tool schemas: hand-authored OpenAI function-calling JSON schemas, original work.
• Conversation content: synthetic; no scraping of any external website, document, or API.
• Languages: English only in this release. Multilingual extensions are planned.

License

• Non-commercial use: Apache License 2.0.
• Commercial use: requires a separate written license. Email contact.tahirrasool@gmail.com.

Citation

@misc{agentforge_multiturn_toolcall_5k,
  title  = {AgentForge-MultiTurn-ToolCall-5k: A Synthetic Multi-Turn Agentic Tool-Calling Dataset with Error Recovery},
  author = {AgentForge},
  year   = {2026},
  note   = {Gated dataset; commercial use requires written license.}
}

Release notes

• v1.0.0 (2026-06-29): initial release. 5,000 conversations, 8 domains, 30.5 % recovery rate.

Contact

Commercial licensing, custom extensions (50k-scale, multilingual, DPO preference pairs, custom domains), and consulting: contact.tahirrasool@gmail.com