GSM8K-700M-v1.2

GSM8K Dense LLM v1.2 (700M Parameters)

A custom Dense LLM architecture pretrained for mathematical reasoning tasks, particularly GSM8K-style problems.

Model Highlights

What's New in v1.2

This release maintains the mastery-level performance of v1.1 while adding additional edge case coverage for robustness. The model now handles a broader range of problem variations, ambiguous wording, and numerical edge cases more reliably. Training included targeted synthetic examples of:

• Problems with awkward phrasing or unconventional wording
• Edge cases with very small/large numbers
• Scenarios requiring careful unit conversions
• Problems with extraneous information
• Boundary conditions in rate/time/distance problems

The result is a more robust model that maintains high accuracy even when problems deviate from standard formulations.

---

Benchmark Results

GSM8K Proficiency Evaluation

The model was evaluated on a 10-question graded assessment spanning difficulty levels 1-10, testing progressive mathematical reasoning skills.

Overall Score: 11/12 (92%)
Proficiency Level: 9-10 (Advanced to Mastery)

---

Example Solutions

The following examples demonstrate the model's step-by-step reasoning capabilities on mastery-level problems.

Example 1: Mixture Problem (Level 10A - Mastery) ✅

Problem:

A coffee shop mixes two types of beans. Type A costs $12 per pound and Type B costs $18 per pound. The shop creates a 50-pound mixture that sells for $14.40 per pound. How many pounds of each type of bean are in the mixture?

Model Output:

Step 1: Let x be the pounds of Type A beans and y be the pounds of Type B beans.
Step 2: Total weight equation: x + y = 50.
Step 3: Cost equation: 12x + 18y = 14.40 × 50 = 720.
Step 4: Simplify cost equation: divide both sides by 6 → 2x + 3y = 120.
Step 5: Solve the system of equations:
  x + y = 50
  2x + 3y = 120
Subtract the first from the second: x = 30.
Then y = 20.
Step 6: Verify total cost: 12 × 30 + 18 × 20 = 360 + 360 = 720, which matches the given total.
Therefore, the mixture contains 30 pounds of Type A beans and 20 pounds of Type B beans.
#### 30

✅ Correct Answer: 30 pounds Type A, 20 pounds Type B

Skills Demonstrated:

• Variable setup with clear definitions
• System of equations construction
• Algebraic simplification (dividing by 6 for cleaner numbers)
• Correct solution method
• Self-verification step

---

Example 2: Round-Trip Rate Problem (Level 10C - Mastery) ✅

Problem:

Emma drives from City A to City B at 60 mph. She immediately returns from City B to City A at 40 mph. If the entire round trip took 5 hours, what is the distance between City A and City B?

Model Output:

1. Let d be the distance between cities in miles.
2. Time to go from A to B: d/60 hours.
3. Time to return from B to A: d/40 hours.
4. Total time: d/60 + d/40 = 5.
5. Find a common denominator (120): 2d/120 + 3d/120 = 5.
6. Combine like terms: 5d/120 = d/24.
7. So, d/24 = 5.
8. Multiply both sides by 24: d = 5 × 24 = 120.
Thus, the distance between cities is 120 miles.
#### 120

✅ Correct Answer: 120 miles

Skills Demonstrated:

• Correct rate/time/distance setup (time = distance/speed)
• Proper fraction arithmetic with common denominators
• Avoided the "average speed trap" (did NOT incorrectly average 60 and 40 to get 50 mph)
• Clean algebraic manipulation
• Correct final answer

⚠️ Note on the "Average Speed Trap": A common failure mode where models incorrectly compute: "Average speed = (60+40)/2 = 50 mph, so distance = 50 × 2.5 = 125 miles". This model correctly recognized that speeds cannot be directly averaged for equal distances—demonstrating true mathematical reasoning rather than pattern matching.

---

Demonstrated Capabilities

Key Findings

• True reasoning ability: Model correctly avoids common mathematical traps (e.g., incorrectly averaging speeds for round-trip problems)
• Enhanced robustness: v1.2 handles edge cases and non-standard problem formulations more reliably
• Clean solution formatting: Produces well-structured step-by-step solutions with verification steps
• Identified training gap: Narrow gap on "combined work-rate with subtraction" problem archetype (e.g., "Three workers together complete a task in X days, Worker A alone takes Y days...")—this appears to be a data coverage issue rather than a reasoning limitation

---

Architecture

• Custom DenseLLM with Grouped-Query Attention (GQA)
• RoPE positional embeddings
• SwiGLU activation
• RMSNorm normalization
• Weight tying (embedding & output)

Technical Specifications

---

Training

This model was pretrained (not fine-tuned) on mathematical reasoning data including GSM8K-style problems. It performs text completion rather than instruction-following.

---

Usage

This model uses a custom architecture. See the repository for loading code.

Loading the Model

import torch
from transformers import AutoTokenizer

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("AlgoDriveAI/GSM8K-700M-v1.2")

# Load model (requires the modeling_dense_llm.py file)
from modeling_dense_llm import load_model
model = load_model("pytorch_model.bin", "config.json", device="cuda")

# Or manually:
from modeling_dense_llm import DenseLLM
import json

with open("config.json") as f:
    config = json.load(f)

model = DenseLLM(
    vocab_size=config["vocab_size"],
    d_model=config["d_model"],
    n_layers=config["n_layers"],
    n_heads=config["n_heads"],
    n_kv_heads=config["n_kv_heads"],
    ff_hidden_mult=config["ff_hidden_mult"],
    qk_norm=config["qk_norm"],
    parallel_residual=config["parallel_residual"],
    max_seq_len=config["max_seq_len"],
)

state_dict = torch.load("pytorch_model.bin", map_location="cuda")
model.load_state_dict(state_dict)
model = model.cuda().bfloat16().eval()

Generation Example

# Math problem completion
prompt = """Question: Sarah has 5 apples. She gives 2 to her friend and then buys 3 more. How many apples does Sarah have now?
Let's solve this step by step:"""

input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].cuda()

output = model.generate(
    input_ids,
    max_new_tokens=256,
    temperature=0.7,
    top_k=50,
    top_p=0.9,
    repetition_penalty=1.1,
    eos_token_id=tokenizer.eos_token_id,
)

print(tokenizer.decode(output[0]))

---

Intended Use

• Mathematical reasoning and problem solving
• Text completion for math-related content
• Research on efficient LLM architectures
• Benchmarking small-scale math reasoning models

Limitations

• This is a pretrained model, not instruction-tuned
• Best used for text completion, not chat/QA
• May require fine-tuning for specific downstream tasks
• Known gap: Combined work-rate problems involving subtraction of reciprocal rates (fixable with targeted training data)

---

Requirements

torch>=2.0
transformers
einops

---

Citation

If you use this model, please cite:

@misc{gsm8k-densellm-700m-v1.2,
  author = {AlgoDriveAI, Christopher Smith},
  title = {GSM8K Dense LLM 700M v1.2},
  email = {[email protected]},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/AlgoDriveAI/GSM8K-700M-v1.2}
}