Modelfile Standards

Intermediate

Enforce consistent Ollama Modelfile structure with required sections, parameter documentation, proper system prompts, and template configuration for reliable custom model builds.

File Patterns

**/Modelfile***/modelfile*

This rule applies to files matching the patterns above.

Rule Content

rule-content.md

# Modelfile Standards

## Rule
All Ollama Modelfiles MUST include a FROM directive, documented SYSTEM prompt, explicitly set temperature, num_ctx, and appropriate stop tokens.

## Required Structure
```dockerfile
# Model: descriptive-name
# Purpose: What this model variant is for
# Base: model-name:tag

FROM model-name:tag

# System prompt — defines model behavior
SYSTEM """
Clear, focused system prompt describing the model's role
and behavioral constraints.
"""

# Inference parameters — documented
PARAMETER temperature 0.2      # Low for code, higher for creative
PARAMETER top_p 0.9            # Nucleus sampling threshold
PARAMETER num_ctx 8192         # Context window size
PARAMETER num_predict 2048     # Max output tokens
PARAMETER stop "<|eot_id|>"    # Stop token for model architecture
```

## Naming Convention
```bash
# Model names: lowercase, descriptive, version tagged
# Format: purpose-language:version
ollama create ts-coder:v1 -f Modelfile.typescript
ollama create python-reviewer:v2 -f Modelfile.python-review
ollama create sql-helper:v1 -f Modelfile.sql
```

### Good Names
```
ts-coder:v1
react-assistant:v2
sql-optimizer:v1
python-debugger:v1
```

### Bad Names
```
my-model          # No purpose, no version
test              # Meaningless
asdf:latest       # Not descriptive
```

## Parameter Rules
| Parameter | Code Gen | Code Review | Creative |
|-----------|----------|-------------|----------|
| temperature | 0.1-0.3 | 0.3-0.5 | 0.7-0.9 |
| num_ctx | 8192+ | 8192+ | 4096 |
| num_predict | 2048 | 4096 | 2048 |
| top_p | 0.9 | 0.95 | 0.95 |

## Anti-Patterns
- Modelfile without SYSTEM prompt (generic behavior)
- No temperature set (defaults vary by model)
- Missing stop tokens (generation runs past intended end)
- No comments explaining parameter choices
- Using :latest tag in FROM (breaks reproducibility)

FAQ

Discussion

Loading comments...

# Modelfile Standards ## Rule All Ollama Modelfiles MUST include a FROM directive, documented SYSTEM prompt, explicitly set temperature, num_ctx, and appropriate stop tokens. ## Required Structure ```dockerfile # Model: descriptive-name # Purpose: What this model variant is for # Base: model-name:tag FROM model-name:tag # System prompt — defines model behavior SYSTEM """ Clear, focused system prompt describing the model's role and behavioral constraints. """ # Inference parameters — documented PARAMETER temperature 0.2 # Low for code, higher for creative PARAMETER top_p 0.9 # Nucleus sampling threshold PARAMETER num_ctx 8192 # Context window size PARAMETER num_predict 2048 # Max output tokens PARAMETER stop "<|eot_id|>" # Stop token for model architecture ``` ## Naming Convention ```bash # Model names: lowercase, descriptive, version tagged # Format: purpose-language:version ollama create ts-coder:v1 -f Modelfile.typescript ollama create python-reviewer:v2 -f Modelfile.python-review ollama create sql-helper:v1 -f Modelfile.sql ``` ### Good Names ``` ts-coder:v1 react-assistant:v2 sql-optimizer:v1 python-debugger:v1 ``` ### Bad Names ``` my-model # No purpose, no version test # Meaningless asdf:latest # Not descriptive ``` ## Parameter Rules | Parameter | Code Gen | Code Review | Creative | |-----------|----------|-------------|----------| | temperature | 0.1-0.3 | 0.3-0.5 | 0.7-0.9 | | num_ctx | 8192+ | 8192+ | 4096 | | num_predict | 2048 | 4096 | 2048 | | top_p | 0.9 | 0.95 | 0.95 | ## Anti-Patterns - Modelfile without SYSTEM prompt (generic behavior) - No temperature set (defaults vary by model) - Missing stop tokens (generation runs past intended end) - No comments explaining parameter choices - Using :latest tag in FROM (breaks reproducibility)