QTM7-4b-1771

QTM7-4B

QTM7-4B is a proof-of-concept math & code reasoning model, trained briefly from Qwen/Qwen3-4B-Base.
It was finetuned for ~4 hours on a single A100 GPU, using lightweight datasets focused on mathematical reasoning and structured problem solving.
This project demonstrates what can be achieved on minimal compute/budget (≈$20 total cost).

---

UPDATE: Observed Performance Shift

This model was explicitly trained using math and code datasets with the intent of achieving higher performance in structured reasoning compared to the base Qwen3-4B model. While quantitative GSM8K metrics show improved math ability, recent qualitative testing suggests an unexpected side effect:

QTM7-4B exhibits significantly enhanced performance in creative writing, narrative generation, and descriptive tasks compared to the Qwen3-4B base model.

The model appears to have utilized the focused finetuning to better understand complex instruction following and structure, which has translated into a superior ability to generate cohesive and evocative creative content.

---

Model Details

• Developed by: Independent researcher (solo project)
• Funding: Self-funded (~$20 total compute cost)
• Model type: Decoder-only transformer for text generation
• Language(s): English
• License: Apache-2.0
• Finetuned from: Qwen/Qwen3-4B-Base

Sources

• Repository: Ma7ee7/QTM7-4b-2hr-checkpoint

---

Uses

Direct Use

• Research into math & code reasoning
• Proof-of-concept for low-budget finetuning on large language models
• New Focus: Evaluation of low-resource impact on creative writing and narrative coherence.

Downstream Use

• Potential basis for math problem solvers or code reasoning assistants
• Experiments in lightweight alignment or evaluation pipelines

Out-of-Scope

• Not suitable for safety-critical, legal, or medical applications
• Not RLHF-aligned; outputs may be unfiltered or ungrounded

---

Bias, Risks, and Limitations

• Inherits biases from Qwen3-4B-Base
• Untested on broader NLP benchmarks (MMLU, ARC, etc.)
• Training was short (~2 hours net, ~4 GPU hours total), so coverage is shallow
• General conversational ability remains base-model level

Recommendation: Treat outputs as experimental. Do not deploy in production or decision-making contexts.

---

Training Details

Training Data

• unsloth/OpenMathReasoning-mini — math reasoning dataset
• nvidia/OpenCodeReasoning — code reasoning tasks
• No GSM8K contamination was found in either the training or post-training data.

Procedure

• Mixed precision: fp16
• Optimizer: AdamW (standard defaults)
• Duration: ~4 hours on 1x NVIDIA A100
• Checkpoint size: ~16 GB (fp16)

---

Evaluation

Setup

• Compared against Qwen/Qwen3-4B (post-trained version)
• Dataset: GSM8K test split (subset of 300 “hard” problems)
• Metrics: Exact match on final numeric answer

Results

Training Loss Curve
Stable convergence toward ~0.63 by step 1750, even as difficulty increased.

---

GSM8K Accuracy (Sampled)
QTM7-4B* scored ~80.7% vs Qwen3-4B’s ~28.0%.

---

Head-to-Head Outcomes
QTM7-4B* won most direct comparisons.

• Only QTM7-4B* correct → 171
• Both correct → 71
• Both wrong → 45
• Only Qwen correct → 13

---

Outcome Breakdown by Model (GSM8K subset)
Side-by-side percentages for correctness vs error types.

• QTM7-4B*: 80.7% correct, 7.3% mismatch, 12.0% truncated
• Qwen3-4B: 28.0% correct, 72.0% mismatch, 0% truncated

---

* QTM7-4B = 2hr checkpoint

---

Environmental Impact

Estimated using MLCO2 Impact Calculator:

• Hardware: NVIDIA A100 (80GB)
• GPU hours: ~4
• Cloud Provider: Google Colab (us-central assumed)
• Carbon emitted: ≈ 1.2 kg CO2eq

(About the same as driving ~5 km in a gasoline car.)

---

Technical Specifications

• Architecture: Qwen3-4B transformer (4B params, decoder-only, rotary embeddings, SwiGLU, grouped query attention)
• Objective: Causal LM finetuning on reasoning tasks
• Software: PyTorch + Hugging Face Transformers + Datasets

---

Summary

QTM7-4B is a minimal-budget proof-of-concept showing that:

• Small compute can still move the needle on reasoning with focused datasets.
• Math reasoning gains were observed even with short finetunes.
• The model is not benchmarked broadly, but shows promise as a low-resource experiment.