aiacontext/mini-enedina-dataset

Mini-Enedina Dataset: Physically Validated Timoshenko Shaft Analysis (60k)

Training dataset for Mini-Enedina 37.5M -- a monotropic language model deliberately small and intensively specialized for structural shaft analysis according to Timoshenko beam theory.

Dataset Description

60,000 synthetic conversations in Harmony-Enedina format (a ChatML variant), covering three progressively complex levels of shaft analysis:

Level	Analysis Scope	Samples	Avg. Tokens/Sample
Bachelor	Deflection (V, M, w, theta)	20,000	8,394
Master	+ Von Mises stress (sigma_VM, sigma_adm)	20,000	9,883
Doctor	+ Fatigue (Marin factors, Goodman criterion)	20,000	12,775

Total: 621M tokens (custom 8,012-token BPE vocabulary).

Every sample is physically validated -- the numerical solver runs to completion, and deflection, stress, and fatigue checks are verified before inclusion in the dataset.

Conversation Format: Harmony-Enedina

Each sample contains a complete conversation with two semantic channels (reasoning + response) and metacognition via headers:

<|system|>
You are Enedina, a structural engineer...
<|end|>
<|user|>
Analyze the transmission shaft with the following parameters...
<|end|>
<|assistant|>
<|ch1_start|>thinking<|ch1_end|>
[Detailed technical reasoning]
<|ch2_start|>response<|ch2_end|>
[Formatted response with LaTeX equations and Python solver code]
<|end|>
<|return|>

Domain-specific tokens (<|shaft|>, <|python|>, <|numerical|>, <|latex|>) demarcate semantic boundaries within the output. The training language is Brazilian Portuguese.

Features (36 columns)

Identification

Column	Type	Description
id	string	Unique SHA-256 hash
problem_type	string	Always shaft_timoshenko
problem_class	string	Problem class
level	string	bachelor, master, or doctor
curriculum_stage	int32	Curriculum stage (0, 1, 2)
analysis_depth	string	deflection, stress, or fatigue

Shaft Parameters

Column	Type	Description
D_ext_mm	float64	External diameter [mm]
D_int_mm	float64	Internal diameter [mm] (0 for solid)
L_mm	float64	Shaft length [mm]
section_type	string	macico (solid) or vazado (hollow)
L_D_ratio	float64	L/D ratio (slenderness)
material_name	string	Material name (e.g., AISI 1045)
E_GPa	float64	Young's modulus [GPa]
S_y_MPa	float64	Yield strength [MPa]
S_ut_MPa	float64	Ultimate tensile strength [MPa]

Loading

Column	Type	Description
n_loads	int32	Number of applied loads
n_bearings	int32	Number of bearings
total_load_kN	float64	Total load [kN]

Analysis Results

Column	Type	Description
V_max_kN	float64	Maximum shear force [kN]
M_max_kNmm	float64	Maximum bending moment [kN.mm]
w_max_mm	float64	Maximum deflection [mm]
w_limit_mm	float64	Deflection limit [mm]

Validation

Column	Type	Description
validation_passed	bool	All validations passed
deflection_passed	bool	Deflection within limit
stress_passed	string	Stress analysis result
fatigue_passed	string	Fatigue analysis result
difficulty_score	float64	Difficulty score [0-1]

Conversation

Column	Type	Description
conversation	string	Full conversation in Harmony format
user_tokens	int32	User prompt token count
assistant_tokens	int32	Assistant response token count
total_tokens	int32	Total conversation token count

Metadata

Column	Type	Description
expected_route	string	Expected route (bachelor/master/doctor)
split	string	Split: train, val, or test
fold	int32	Cross-validation fold
solve_time_ms	float64	Solver execution time [ms]
template_variation	int32	Template variation used

Token Statistics

Metric	Value
Total tokens	621,028,968
Mean per sample	10,350
Minimum	8,280
Maximum	12,996
Median	9,881
User tokens (mean)	358
Assistant tokens (mean)	9,932

Usage

from datasets import load_dataset

dataset = load_dataset("aiacontext/mini-enedina-dataset")

# Access splits
train = dataset["train"]       # 48,000 samples
val = dataset["validation"]    # 6,000 samples
test = dataset["test"]         # 6,000 samples

# Filter by level
bachelor = train.filter(lambda x: x["level"] == "bachelor")
doctor = train.filter(lambda x: x["level"] == "doctor")

# Access a conversation
sample = train[0]
print(sample["conversation"][:500])
print(f"Tokens: {sample['total_tokens']}")
print(f"Level: {sample['level']}")

Data Generation

The data was generated by Factorium SciML -- a Julia/Python framework for scientific synthetic data generation:

1. Numerical solver (Python/NumPy): solves each Timoshenko problem with randomized parameters
2. Variable templates: 97k description combinations in PT-BR (46^3 variations)
3. Numerical grounding: a citation system that ensures consistency between input parameters and values in the response
4. Validation: every sample is verified (deflection, stress, fatigue) before inclusion

Governing Equations (Timoshenko with self-weight)

V(x)  = RA - q*x - sum(Fi * H(x - xi))           (Shear force)
M(x)  = RA*x - q*x^2/2 - sum(Fi*(x-xi)*H(x-xi)) (Bending moment)
dw/dx = theta + V/(kappa*G*A)                     (Deflection with shear)
q     = rho * g * A * 1e-9                        (Distributed self-weight [N/mm])

Trained Model

The model trained on this dataset is available at: aiacontext/mini-enedina

Parameter	Value
Architecture	Dense Transformer (RoPE, RMSNorm, SwiGLU)
Parameters	37.5M
Framework	MLX (Apple Silicon)
Test Loss	0.0787
Test Perplexity	1.08

Citation

If you use this dataset, please cite:

@article{leitaofilho2026minienedina,
  title={Mini-Enedina: A Domain-Specialized Small Language Model for Structural Shaft Analysis Using Timoshenko Beam Theory},
  author={Leit{\~a}o Filho, Antonio de Sousa and Barros Filho, Allan Kardec Duailibe and Lima, Fabr{\'i}cio Saul and Santos, Selby Mykael Lima dos and Sousa, Rejani Bandeira Vieira},
  year={2026}
}

Acknowledgments

This work was supported by Aia Context Ltda. and by FINEP -- Funding Authority for Studies and Projects, a Brazilian government agency for science, technology, and innovation linked to the Ministry of Science, Technology and Innovation (MCTI), under Contract No. 03.25.0080.00.

License

CC-BY-4.0