SII-Enigma

Davinci-Tool-30B

README

License: apache-2.0

Method

Davinci-ToolData is built by two complementary pipelines that share one execution and verification interface:

EGTS — Execution-Grounded Task Synthesis (authored systems). An explorer interacts with a resettable executable world; user tasks are synthesized from the recorded outcome; the world is reset and the task is replayed with a guide, and the replay is accepted only if it recovers the explored outcome. Structural checks, a hard factual judge, and a soft quality judge filter the results.
EES — Executable Environment Synthesis (synthesized systems). Audited executable domains (191 domains in 19 categories) produce state-grounded tasks with executable verifiers within one domain (EES-Single), and are composed into shared multi-domain worlds (246 compositions) with shared entities, linking fields, cross-domain policies, and coherence constraints (EES-Cross).

Training has two stages: SFT on the verified demonstrations, then decision-level DPO, which reuses restorable states and executable verifiers to fork the SFT policy at decision points, execute alternative actions from the same state, and keep pairs whose verified downstream outcomes differ (1,150 pairs).

Main results

Table with columns: Model, τ²-bench Retail, τ²-bench Telecom, τ²-bench Avg., BFCL v4 MT Avg., MCP-Atlas
Model	τ²-bench Retail	τ²-bench Telecom	τ²-bench Avg.	BFCL v4 MT Avg.	MCP-Atlas
Qwen3-Coder-480B-A35B-Instruct	77.5	39.5	58.5	43.5	28.1
Kimi-K2-Instruct	70.6	65.8	68.2	33.4	59.6
MiniMax-M2.1	81.6	87.0	84.3	71.4	59.6
GLM-4.7	79.8	98.2	89.0	65.4

All scores are from the authors' evaluations under the official protocols; MCP-Atlas uses the 89 tasks that do not require external API keys. Under the same backbone and SFT hyperparameters, Davinci-ToolData exceeds the public Toucan-1.5M release by 7.4 / 18.7 / 30.3 points.

Transfer to smaller models (SFT on Davinci-ToolData):

Table with columns: Model, τ²-bench Avg., BFCL v4 MT Avg.
Model	τ²-bench Avg.	BFCL v4 MT Avg.
Qwen3-8B → + Davinci-ToolData	32.5 → 74.6	15.4 → 33.9
Qwen3-14B → + Davinci-ToolData	33.4 → 76.8	21.1 → 41.4

Data sources are complementary (SFT on Qwen3-Coder-30B-A3B-Instruct):

Table with columns: EGTS, EES-Single, EES-Cross, τ²-bench, BFCL v4, MCP-Atlas
EGTS	EES-Single	EES-Cross	τ²-bench	BFCL v4	MCP-Atlas
✓			78.9	57.4	28.1
	✓		82.0	54.5	29.2
		✓

Quickstart

The model uses the Qwen3-Coder chat template and tool-call format (chat_template.jinja).

vLLM

bash

vllm serve SII-Enigma/Davinci-Tool-30B \  --enable-auto-tool-choice \  --tool-call-parser qwen3_coder \  --max-model-len 65536

Then call it through any OpenAI-compatible client with a tools list.

Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "SII-Enigma/Davinci-Tool-30B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto") tools = [{    "type": "function",    "function": {        "name": "get_weather",        "description": "Get the current weather for a city.",        "parameters": {            "type": "object",            "properties": {"city": {"type": "string"}},            "required": ["city"],        },    },}]messages = [{"role": "user", "content": "What's the weather in Shanghai?"}] inputs = tokenizer.apply_chat_template(    messages, tools=tools, add_generation_prompt=True, return_tensors="pt").to(model.device)outputs = model.generate(inputs, max_new_tokens=512, use_cache=True)print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=False))

generation_config.json ships the Qwen3-Coder default sampling settings (temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.05).

Model details

Table

Architecture	`Qwen3MoeForCausalLM`
Parameters	30.5B total, 3.3B activated
Layers / experts	48 layers, 128 experts, 8 activated per token
Attention	32 query heads, 4 KV heads (GQA)
Context (config)	262,144 tokens
Precision	BF16 safetensors

Training settings

Table with columns: Setting, SFT, DPO
Setting	SFT	DPO
Initialization	Qwen3-Coder-30B-A3B-Instruct	SFT checkpoint (also the frozen reference)
Data	Davinci-ToolData, 30,395 trajectories	1,150 decision-level preference pairs
Learning rate / schedule	1e-5, cosine, no warmup	5e-6, cosine, 3% warmup
Epochs	4	1
Max sequence length	32,768	24,576
Batch per device × grad. accum.

License

Released under the Apache 2.0 License, following the base model Qwen3-Coder-30B-A3B-Instruct.

Citation

bibtex

@article{yuan2026davincitool,  title   = {Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments},  author  = {Yuan, Xiaoyang and Zhao, Ji and Hu, Xingyu and Wang, Wenhan and Hu, Yutao and Bin, Yi and Shao, Wenqi and Zhang, Qiaosheng and Liu, Pengfei and Qiao, Yu},  journal = {arXiv preprint arXiv:XXXX.XXXXX},  year    = {2026}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

SII-Enigma

Model Tree

Base

Qwen/Qwen3-Coder-30B-A3B-Instruct

Fine-tuned

this model

Input Modalities

Text

Output Modalities