ANMOLGPT-4B-v0.5

ANMOLGPT-4B-v0.5

ANMOLGPT-4B-v0.5 is the fifth experimental release of the ANMOLGPT family of small language models.

This version is based on Qwen3.5-4B and was further fine-tuned with a focus on mathematical reasoning using GSM8K.

The primary objective of v0.5 was to investigate whether targeted mathematical fine-tuning could improve reasoning performance while preserving the model's broader capabilities.

Model Details

• Model: ANMOLGPT-4B-v0.5
• Base model: Qwen/Qwen3.5-4B
• Parameters: ~4B
• Architecture: Causal Language Model
• Language: English
• Primary task: Text generation and reasoning
• Training focus: Mathematical reasoning
• Fine-tuning dataset: GSM8K
• Framework: Transformers / Unsloth
• Quantization during training: 4-bit
• Fine-tuning approach: Parameter-efficient fine-tuning

---

What's New in v0.5?

The main change in v0.5 is targeted training on GSM8K to improve mathematical reasoning.

Compared with ANMOLGPT-4B-v0.4, v0.5 shows a substantial improvement on GSM8K while largely preserving performance across several general-purpose benchmarks.

Key GSM8K Results

The improvement in strict-match accuracy is particularly notable, suggesting that the model became better at producing answers in the expected mathematical format.

---

Benchmark Results

All evaluations below were performed using lm-evaluation-harness.

GSM8K

5-shot evaluation.

---

HellaSwag

0-shot evaluation.

---

PIQA

0-shot evaluation.

---

ARC-Easy

0-shot evaluation.

---

Winogrande

0-shot evaluation.

---

TruthfulQA

0-shot evaluation.

Note: TruthfulQA decreased compared with v0.4. This is an observed trade-off following the GSM8K-focused fine-tuning and is reported transparently.

---

MMLU

0-shot evaluation.

Overall

73.78% ± 0.35%

Selected Mathematical Subjects

---

v0.4 → v0.5 Comparison

---

Interpretation

The results suggest that GSM8K-focused fine-tuning substantially improved mathematical reasoning without causing a broad degradation across the evaluated general-purpose benchmarks.

The largest improvement was observed on GSM8K:

54.36% → 60.42%

while MMLU, ARC-Easy, PIQA, HellaSwag and Winogrande remained broadly stable or improved.

However, TruthfulQA decreased:

54.61% → 49.01%

This highlights an important trade-off in targeted fine-tuning of smaller language models: improving a specific capability can affect other capabilities.

Future versions of ANMOLGPT will investigate methods for improving mathematical reasoning while preserving factuality and general reasoning performance.

---

Intended Use

ANMOLGPT-4B-v0.5 is intended primarily for:

• Research into small language models
• Mathematical reasoning experiments
• Instruction-following research
• Local inference experimentation
• Model fine-tuning experiments
• Benchmarking and evaluation
• Educational experimentation

It can also be used as a general-purpose text-generation model, although it should not be considered a replacement for significantly larger frontier models.

---

Limitations

ANMOLGPT-4B-v0.5 is an experimental research model.

Known limitations include:

• Mathematical reasoning remains imperfect.
• The model may produce incorrect calculations or reasoning.
• Truthfulness performance decreased after GSM8K-focused training.
• The model can hallucinate information.
• Benchmark performance does not necessarily translate directly into real-world reliability.
• The model has substantially fewer parameters than modern frontier models.
• No claim is made that the model is suitable for safety-critical or high-stakes applications.

Outputs should therefore be independently verified, especially for financial, medical, legal, scientific or other high-impact applications.

---

Training Approach

The model was developed through iterative experimentation.

The ANMOLGPT development process has focused on understanding how targeted datasets affect the capabilities of a relatively small language model.

v0.5 specifically investigates:

Can targeted mathematical reasoning training significantly improve a ~4B parameter model without causing widespread capability degradation?

The results indicate that the answer is partially yes.

---

Evaluation

Evaluation was performed using:

EleutherAI LM Evaluation Harness

The model was evaluated using the following tasks:

• GSM8K
• MMLU
• ARC-Easy
• PIQA
• HellaSwag
• Winogrande
• TruthfulQA

Unless otherwise specified, evaluations used 0-shot settings.

GSM8K was evaluated using 5-shot evaluation.

---

Example Usage

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "anmoldhandhania93/ANMOLGPT-4B-v0.5"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

prompt = "Solve this step by step: If a train travels 60 km in 1.5 hours, what is its average speed?"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))