prxshetty/llama-3.1-8b-story-point-estimator

Llama 3.1 8B Story Point Estimator

A fine-tuned Llama 3.1 8B model specialized for agile story point estimation in software development workflows.

---

Overview

Item	Details
Base checkpoint	unsloth/Meta-Llama-3.1-8B
Fine-tune method	LoRA (PEFT) with Unsloth
Training run	1 epoch • 60 max steps • Custom dataset
Trainable params	LoRA adapters only
Task	Story point estimation (1-20 scale)
Hardware	Google Colab (T4/V100)
License	Llama 3.1
Intended use	Agile development & estimation assistance

---

Usage

from unsloth import FastLanguageModel

Load the model
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="prxshetty/llama-3.1-8b-story-point-estimator",
max_seq_length=2048,
dtype=None,
load_in_4bit=True,
)

Switch to inference mode
FastLanguageModel.for_inference(model)

Format your prompt using Llama 3.1 chat template
prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>

You are an expert agile estimation assistant. Your task is to estimate story points for a software development task.

CONTEXT:
Story points are a team-specific, unit-less measure of relative effort required to complete a backlog item in agile software development.

ESTIMATION SCALE:
Range: minimum 1, maximum 20

TRAINING EXAMPLE:
Issue Title: {title}
Issue Description: {description}

INSTRUCTIONS:
Based on the example above, analyze the complexity, technical difficulty, unknowns, and scope of work.
Provide only the story point estimate as a single integer.<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=False)
response = tokenizer.decode(outputs[inputs['input_ids'].shape:], skip_special_tokens=True)

---

Training Details

Parameter	Value
Epochs	1
Max Steps	60
Batch Size	2 (per device)
Gradient Accumulation	4 steps
Learning Rate	2e-4
Optimizer	AdamW 8-bit
Weight Decay	0.01
Warmup Steps	5

The model was trained using train_on_responses_only to optimize only on the story point outputs, not the input prompts.

---

Responsible Use

Intended for estimation assistance only. This model provides suggestions based on training patterns and should not replace human judgment in agile planning. Always:

• Use as a starting point for team discussions
• Validate estimates with domain experts
• Consider team-specific velocity and context
• Review and adjust based on historical performance

The author and base-model creators accept no liability for project planning decisions based on this model's outputs.

---

Model Architecture

Built on Meta's Llama 3.1 8B architecture with LoRA (Low-Rank Adaptation) fine-tuning for parameter efficiency. The model uses the standard Llama 3.1 chat template format for consistent inference.

---