agentic-think-v1

Agentic tool-calling SFT corpus with teacher-generated <think> reasoning on every assistant round. 212,396 training rows across 9 sources, built by ENERZAi for small-model (1.7B ternary) agentic SFT.

Each row is a full multi-turn conversation: system prompt, user turns, assistant turns (with an inline <think>...</think> block before the visible content / tool call), tool calls and tool results.

Schema

field	type	description
id	str	<source>-<episode> unique id
source	str	one of the 9 sources below
tools	str (JSON)	OpenAI-style function catalog for the episode
messages	str (JSON)	OpenAI-style chat messages; assistant content contains <think>...</think>
history_mode	str, optional	present on robot rows only

from datasets import load_dataset
ds = load_dataset("HBKenerzai/agentic-think-v1", split="train")
# or a single source:
ds = load_dataset("json", data_files={"train": "hf://datasets/HBKenerzai/agentic-think-v1/data/glaive.jsonl"})

Sources & counts

file	rows	upstream	upstream license
nemotron_tc.jsonl	53,751	nvidia/Nemotron (tool-calling split)	CC-BY-4.0
glaive.jsonl	44,980	glaive-function-calling-v2	Apache-2.0
xlam.jsonl	40,000	Salesforce/xlam-function-calling-60k	CC-BY-4.0
synth_apigen.jsonl	28,932	APIGen-style synthetic (ours)	CC-BY-4.0
nemotron_ia.jsonl	21,929	nvidia/Nemotron (interaction split)	CC-BY-4.0
toolace.jsonl	10,322	Team-ACE/ToolACE	Apache-2.0
robot.jsonl	5,872	grid-robot FC synthetic (ours)	CC-BY-4.0
apigen_mt.jsonl	4,778	Salesforce/APIGen-MT	CC-BY-4.0
hermes.jsonl	1,832	NousResearch hermes-function-calling	Apache-2.0

Multi-turn sources (nemotron_*) were filtered to ≥3 user turns (nemotron_tc: 80% 3–5 turns / 20% ≥6; nemotron_ia: 47% / 53%).

Generation method

• Teacher: Qwen3.8-27B (FP8), enable_thinking disabled on the visible channel; think text generated in annotate mode (teacher sees the gold action and writes the reasoning that leads to it), one think block per assistant round.
• Rejection (no-call) episodes: 20,990 rows where the correct behavior is to not call any function. Their single-function catalogs were augmented with 9–15 distractor functions (overlap-guarded against the user request) so the model learns rejection under a realistic catalog, not an empty one.
• Function-catalog augmentation + shuffle applied throughout (distractor injection).

Known limitations

• Annotate-mode think is written after the teacher sees the gold action, so on no-call rounds the reasoning tends to open with the verdict ("No function is called...") — i.e. partially post-hoc rationalization rather than genuine deliberation. A react-style (gold-blind, match-filtered) regeneration is in progress as v2.
• <think> blocks are teacher-generated and not human-verified.

Provenance

Built on the agentic-datagen pipeline (LLM_Hard_STE, feature/agentic-datagen-vllm) by ENERZAi. Used to train the ternary (W1.58) Qwen3-1.7B v23 SFT checkpoint and its think-DPO line.