Trovus

The official repository of the paper, "Reasoning on a Budget: How Teacher Signals Shape Efficiency Frontiers for Small Language Models"

Search for Models

# Interactive search
trovus search

# Example searches:
# - "microsoft deberta"
# - "qwen 270m base"  
# - "phi mini instruct"

Download Models

# Download with research mode (recommended)
trovus download microsoft/deberta-v3-small --output-dir ./models --research-mode

# Download minimal files only

# Custom file patterns
trovus download microsoft/phi-3-mini --include "*.safetensors" "*.json" --exclude "*.bin"

Model Card Management

# Parse a model card for download information
trovus parse-card ./model_cards/microsoft_deberta-v3-small_card.md

# The search interface can save model cards automatically

Cache Management

# List downloaded models
trovus list

# Show cache information
trovus cache-info

# Get detailed model info
trovus info microsoft/deberta-v3-small

# Remove models
trovus remove microsoft/deberta-v3-small

Command Reference

Search Commands

• trovus search - Interactive model search with fuzzy matching
• trovus parse-card <file> - Extract information from model cards

Download Commands

• trovus download <model> - Download models with flexible options
  • --output-dir - Custom download directory
  • --research-mode - Optimized for research (all weights + configs, exclude specialized formats)
  • --minimal - Essential files only (configs + model weights: safetensors/bin/h5/model)
  • --include - File patterns to include
  • --exclude - File patterns to exclude
  • --force - Force re-download

Management Commands

• trovus list - List cached models with sizes and dates
• trovus info <model> - Detailed information about a specific model
• trovus cache-info - Overall cache statistics
• trovus remove <model> - Remove models from cache

Evaluate Commands

• trovus evaluate <model> --method <sft|cot-d|rl> - Run teacher-signal evaluation flows
  • SFT (implemented): launches supervised fine-tuning with LoRA/TRL on a registered dataset
  • CoT-D / RL (stubs): records config and prepares output directories for upcoming pipelines
  • Key flags:
    • --dataset (default gsm8k)
    • --epochs, --learning-rate, --per-device-train-batch-size, --gradient-accumulation-steps
    • --use-4bit for 4-bit quantization, --lora-rank, --target-modules
    • --output-dir for run artifacts, --cache-dir for HF cache overrides

Download Modes

Research Mode (--research-mode)

Downloads all model weights and configurations while excluding specialized formats:

• ✅ Includes: *.safetensors, *.bin, *.h5, *.model, *.json, *.txt, *.py, *.md
• ❌ Excludes: *.msgpack, *.onnx, *.tflite, *.gguf, framework-specific duplicates

Minimal Mode (--minimal)

Downloads only essential files needed to use the model:

• ✅ Includes: Config files + all model weight formats (*.safetensors, *.bin, *.h5, *.model)
• ❌ Excludes: Less common formats (*.msgpack, *.onnx, *.tflite)

Examples

# Search and download workflow
trovus search
# Type: "microsoft deberta"
# Save model card when prompted
trovus parse-card ./model_cards/microsoft_deberta-v3-small_card.md
trovus download microsoft/deberta-v3-small --output-dir ./models --research-mode

# Quick download for inference
trovus download Qwen/Qwen3-0.6B --minimal --output-dir ./models

# Fine-tune a locally cached model with SFT (LoRA defaults)
trovus evaluate Qwen/Qwen3-0.6B --method sft --dataset gsm8k --epochs 10

# Custom download with specific files
trovus download microsoft/phi-3-mini \
  --include "*.safetensors" "config.json" "tokenizer*" \
  --output-dir ./models

# Check what you have downloaded
trovus list
trovus cache-info

To-do:

• Implement the search retriever + model card downloader.
• Convert "python -m trovus" into a universal command: "using trovus"
• Include ability to download model weights.
• Prepare the dataset for question-answer geneation + pairs.
• Finalize fine-tuning techniques and publish dataset onto HF Hub.
• Implement the fine-tuning metrics which will be the foundation for the efficiency frontiers.
• Have the efficiency frontiers be saved in a reproducible and visualizable format.
• Start with finetuning the easiest technique onto the smallest model for starters.