Charley890/adaption_adaptive_math_2

Overview

Adaptive Math 2

A mathematics-specialized instruction dataset designed to improve reasoning, structured problem solving, and educational AI assistants through supervised fine-tuning with Adaptation Labs AutoScientist.

Research Snapshot

Dataset : https://huggingface.co/datasets/Charley890/adaption-adaptive-math-2

Adaptive Math 2 focuses on educational mathematical reasoning rather than simple answer prediction.

Covers

• —Algebra
• —Geometry
• —Arithmetic
• —Number Theory
• —Statistics
• —Word Problems
• —Mathematical Reasoning
• —Multi-step Solutions

Characteristics

Structured instruction format

Educational explanations

Curriculum-oriented questions

Reasoning-aware responses

Clean supervised fine-tuning format

Example Dataset Samples

Example

Instruction

Solve:
4x - 9 = 19

Expected Response

4x = 28

x = 7

Educational Impact

Adaptive Math 2 is intended for:

• —AI tutors
• —Educational assistants
• —Mathematical reasoning
• —Homework support
• —Classroom demonstrations
• —STEM education
• —Benchmark evaluation

The dataset emphasizes transparent reasoning instead of answer memorization.

Mathematical Training Specification

model: basemodel: "meta-llama/Llama-4-Scout-17B-16E-Instruct" approximatemodelsize: "109B parameters" trainingmethod: "Supervised Fine-Tuning (SFT)" adaptationmethod: "LoRA" dataformat: "Chat"

mathematical_formulation:

objective: description: "The adapted model minimizes the supervised language-modeling loss over the Adaptive Math 2 dataset." equation: | θ* = argmin_θ L(θ)

languagemodelloss: equation: | L(θ) = -Σᵢ log Pθ(yᵢ | xᵢ)

lora: description: "Instead of updating the full model weights, LoRA learns a low-rank update." equation: | W' = W + ΔW ΔW = (α/r)BA

parameters: rankr: 16 alpha: 32 dropout: 0 scalingfactor: | α/r = 32/16 = 2

effective_update: equation: | ΔW = 2BA

optimization: learningrate: 0.00005 weightdecay: 0 maxgradientnorm: 2 optimizer_constraint: | ||g||₂ ≤ 2

trainingschedule: epochs: 5 evaluations: 5 evaluationfrequency: | 5 evaluations / 5 epochs = 1 evaluation per epoch

scheduler: type: "Linear" numcycles: 0.5 warmupratio: 0.03

warmup: equation: | T_warmup = 0.03T

batch: batch_size: "max"

target_modules: count: 10 modules:

• —"k_proj"
• —"o_proj"
• —"q_proj"
• —"v_proj"
• —"shared_expert.gate"
• —"sharedexpert.upproj"
• —"sharedexpert.downproj"
• —"feedforward.gateproj"
• —"feedforward.upproj"
• —"feedforward.downproj"

trainingobjective: equation: | θLoRA* = argmin_{A,B} L(W + (α/r)BA)

interpretation: rank: "r = 16 controls the low-rank adaptation capacity." scaling: "α/r = 2 controls the magnitude of the LoRA update." regularization: "LoRA dropout = 0 and weight decay = 0." stability: "Gradient norm is clipped at 2." schedule: "Learning rate follows a linear schedule after a 3% warmup."

Training Interpretation

benchmark: adaptationstrategy: "Parameter-efficient fine-tuning" objective: "Improve mathematical reasoning while preserving the pretrained model." fullparameterupdate: false lowrank_update: true

key_result: statement: | Adaptive Math 2 applies a low-rank parameter update rather than retraining the complete 109B-parameter model.

mathematicalsummary: | Wadapted = W_base + 2BA

meaning:

• —"W_base represents the pretrained model."
• —"A and B are learned low-rank matrices."
• —"r = 16 defines the adaptation rank."
• —"α = 32 gives a scaling factor of 2."
• —"Only the selected target modules receive LoRA updates."

Reproducibility

configuration: trainingmethod: "SFT" trainingtype: "LoRA" epochs: 5 learningrate: 0.00005 warmupratio: 0.03 weightdecay: 0 maxgradnorm: 2 lorarank: 16 loraalpha: 32 loradropout: 0 scheduler: "linear" schedulercycles: 0.5 evaluations: 5 batchsize: "max"

credit: adaptivedata: "Adaptive Data by Adaption Labs" trainingevaluation: "AutoScientist"

Adaptive Math 2 was developed using the Adaption Lab AutoScientist pipe

📊 Model Performance

[image]

{
  "job_id": "8db3bddd-326c-44ba-8440-2456d10d33f2",
  "training_experiment_id": "78a0fd31-7d13-40cf-bc55-fb2d2bf9e92c",
  "original_model_name": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
  "trained_model_name": "adaption_adaptive_math_2",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 16,
    "n_evals": 5,
    "n_epochs": 5,
    "batch_size": "max",
    "lora_alpha": 32,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.03,
    "weight_decay": 0,
    "learning_rate": 0.00005,
    "max_grad_norm": 2,
    "base_model_size": "109B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "linear",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "k_proj,o_proj,q_proj,v_proj,shared_expert.gate_proj,shared_expert.up_proj,shared_expert.down_proj,feed_forward.gate_proj,feed_forward.up_proj,feed_forward.down_proj"
  }
}

Training Data

The model was trained on 1,306 rows of adapted data with the following domain distribution: math (77%), code (8%), science (8%), academic-education (8%).

Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.

[image]

How to use

pip install torch transformers peft

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
ADAPTER = "<this-repo-id>"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge the LoRA weights into the base for faster inference
model = model.merge_and_unload()
model.eval()

tokenizer = AutoTokenizer.from_pretrained(BASE)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))