A commercial-grade, synthetic, multilingual multi-turn agentic
tool-calling dataset for SFT and DPO post-training. The premium successor
to AgentForge-MultiTurn-ToolCall-5k.
⚠️ 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 on request
Click "Request access" above. Briefly describe your research.
Commercial use (training models you sell, embed in a paid product, internal company use at a company with > $1M ARR or > 10 employees)
Requires a commercial license
Email contact.tahirrasool@gmail.com with subject AgentForge Premium Commercial License. Include: company name, intended use, deployment scale.
Apache-2.0 applies to non-commercial use only. Commercial use without a
signed license is prohibited. If unsure whether your use is commercial, assume
it is and email contact.tahirrasool@gmail.com.
What's new in v2 (vs. v1)
Capability
v1 (5k)
v2 (50k + 5k DPO)
Base conversations
5,000
50,000
Domains
8
12 (added healthcare, legal, hr, cloud-infra)
Languages
English only
8 languages (en, es, fr, de, zh, ja, hi, ar)
Error-recovery rate
30 %
50.5 %
Reasoning traces (chain-of-thought)
—
On every assistant turn
DPO preference pairs
—
5,000 (6 distinct rejection modes)
Difficulty tiers
easy / medium / hard
easy / medium / hard / expert
Code-execution tools
—
Yes (Python sandbox + SSM shell + SQL)
Total tool calls
18,481
195,889
Dataset structure
default config — 50,000 SFT conversations
from datasets import load_dataset
ds = load_dataset("JDKdev/agentforge-premium-v2", token="hf_...")
# ds["train"] → 50,000 records
field
type
description
id
string
Unique id, e.g. afp_00001.
domain
string
One of 12 domains (see coverage below).
language
string
One of en, es, fr, de, zh, ja, hi, ar.
difficulty
string
easy, medium, hard, or expert (= hard + non-English).
includes_recovery
bool
Whether the trajectory recovers from a tool failure.
num_turns
int
Total messages in the conversation.
num_tool_calls
int
Total tool invocations.
tools
list[dict]
OpenAI-compatible function schemas. (JSON-stringified in parquet.)
conversations
list[dict]
ShareGPT-style messages; assistant turns include a reasoning field with chain-of-thought.
dpo config — 5,000 preference pairs
ds = load_dataset("JDKdev/agentforge-premium-v2", "dpo", token="hf_...")
# ds["train"] → 5,000 records with chosen/rejected
field
type
description
id
string
afp_dpo_00001 ... afp_dpo_05000.
domain
string
Inherited from the base record.
language
string
Inherited.
difficulty
string
Inherited.
includes_recovery
bool
All DPO pairs use recovery traces (more interesting preference signal).
rejection_mode
string
One of: wrong_tool, missing_required_arg, hallucinate_success, skip_verification, wrong_param_value, ignore_error.
tools
list[dict]
Function schemas available.
chosen
list[dict]
Correct trajectory (with reasoning).
rejected
list[dict]
Trajectory with a realistic failure injected.
Coverage
By domain
Domain
Records
Unique tools
finance
4,167
5
travel
4,167
5
ecommerce
4,167
6
devops
4,167
5
crm
4,167
5
calendar
4,167
5
email
4,167
5
database
4,167
5
healthcare
4,166
5
legal
4,166
5
hr
4,166
5
cloud_infra
4,166
6
Total
50,000
57 unique
By language
Language
Records
%
en
29,794
59.6 %
es
3,543
7.1 %
zh
3,129
6.3 %
fr
3,074
6.1 %
de
2,976
6.0 %
hi
2,604
5.2 %
ja
2,453
4.9 %
ar
2,427
4.9 %
By difficulty
Tier
Records
Notes
easy
12,422
No recovery, English.
medium
12,319
No recovery, English.
hard
15,057
Includes recovery, English.
expert
10,202
Includes recovery, non-English.
DPO rejection modes (5,000 pairs)
Mode
Pairs
What the rejected response does wrong
wrong_tool
836
Calls an unrelated tool instead of the correct one.
missing_required_arg
814
Omits a required argument.
hallucinate_success
856
Claims success without calling the tool.
skip_verification
813
Skips the verify-then-act step.
wrong_param_value
874
Passes a corrupted parameter value.
ignore_error
807
Proceeds as if a failed tool call succeeded.
Intended use
SFT on 50k base conversations — train small/mid LLMs (1B–14B) to:
decide when to call tools vs. answer from parametric knowledge,
emit OpenAI-style function calls correctly,
chain multi-turn tool sequences,
recover from realistic tool failures,
reason explicitly before each action (chain-of-thought).
DPO / IPO / KTO on 5k preference pairs — sharpen the model's preference for verification, correct tool selection, and honest failure handling.
Multilingual agent evaluation — slice by language to measure non-English agentic capability.
Curriculum learning — order: easy → medium → hard → expert.
Provenance & generation
Generation method: deterministic Python generator with fixed seed (20260629). Fully synthetic; no scraping of any external website, document, or API.
Tool schemas: hand-authored OpenAI function-calling JSON, original work.
Multilingual translations: hand-translated system prompts and policy text for 8 languages. User prompts are kept in English for parser compatibility (industry standard for function-calling datasets).
DPO rejected variants: produced by deterministic mutation of chosen trajectories (6 distinct failure modes), so chosen/rejected differ in a controlled, explainable way.
No PII, no real customer data, no copyrighted material. All names, MRNs, account ids, deal names, etc. are randomly generated.
Reproducibility
The generator script (build_agentforge_premium.py) is deterministic. Running with seed 20260629 reproduces this dataset byte-for-byte (modulo random.shuffle ordering).
@misc{agentforge_premium_v2,
title = {AgentForge-Premium-v2: A 50k Multilingual Multi-Turn Agentic Tool-Calling Dataset with Reasoning Traces and DPO Pairs},
author = {AgentForge},
year = {2026},
note = {Gated dataset; commercial use requires written license.}
}
Release notes
v2.0.0 (2026-06-29): initial premium release. 50k base + 5k DPO, 12 domains, 8 languages, 50.5 % recovery rate, reasoning traces on every assistant turn.
