DiwakarTalwar/Anvi-0.01-test

Anvi 0.01

Anvi is a lightweight QLoRA fine-tuned supervisor model based on Qwen/Qwen2.5-0.5B-Instruct.

The goal of Anvi is not to act as the final conversational assistant. Instead, it analyzes a user's request and produces a structured routing decision that can be passed to a larger AI system.

Anvi 0.01 is an early experimental version of the Anvi supervisor/routing system.

Model Details

Model Description

Anvi is a small language model fine-tuned with QLoRA for request classification and capability routing.

It is trained on examples covering:

• —Programming
• —Debugging
• —Mathematics
• —Weather
• —Database queries
• —Vision
• —OCR
• —Translation
• —Memory retrieval
• —Calendar and task management
• —Email-related tasks
• —Multi-step requests
• —Clarification

The intended architecture is to use Anvi as a lightweight decision-making layer before specialized models and external tools are invoked.

Model Information

• —Model name: Anvi 0.01
• —Model type: QLoRA fine-tuned causal language model / routing supervisor
• —Base model: Qwen/Qwen2.5-0.5B-Instruct
• —Fine-tuning: QLoRA / LoRA
• —Frameworks: PEFT, Transformers, Unsloth
• —Language: English, with some Hinglish-oriented training examples
• —License: Apache 2.0

Intended Use

Direct Use

Anvi is intended to receive a natural-language request and determine which capability should handle it.

A typical routing decision may look like:

{
  "task": "programming",
  "tool": null,
  "model": "coding_expert"
}

The exact routing schema is application-specific and may evolve in future versions.

Downstream Use

Anvi is intended to be integrated into a larger AI system where its output is used to select specialized models, tools, or workflows.

User request
     |
     v
   Anvi
     |
     +----> math_expert
     +----> coding_expert
     +----> vision_expert
     +----> weather tool
     +----> database tool
     +----> memory system
     +----> multi-step workflow

Anvi's primary role is routing and classification rather than direct tool execution.

Out-of-Scope Use

This is an experimental routing model and should not be treated as a fully capable general-purpose assistant or as an authoritative source of factual information.

It should not be relied upon by itself for:

• —Medical, legal, financial, or other high-stakes decisions
• —Autonomous execution of external actions
• —Security-critical routing without additional validation
• —Production deployment without application-level output validation

Limitations

This is an early experimental release.

Known limitations include:

• —The model can occasionally produce malformed or unexpected JSON.
• —Closely related routing categories can be difficult to distinguish.
• —Tool and model selections may not always match the intended routing schema.
• —Multi-step requests can be difficult to classify consistently.
• —The model may still behave like the original instruction-tuned model on some prompts.
• —The training data and routing schema are still being refined.
• —No formal benchmark results are included in this release.

The model should therefore be considered an experimental checkpoint rather than a production-ready router.

Recommendations

For downstream applications, validate the model's output before executing any tool or workflow.

Recommended validation:

1. 1.Parse the generated output as JSON.
2. 2.Validate it against the application's routing schema.
3. 3.Check that selected tasks, tools, and models are allowed.
4. 4.Reject or re-route invalid outputs.
5. 5.Keep actual tool execution outside the model.

How to Get Started

Anvi is distributed as a PEFT/QLoRA adapter. The base model is required to load the adapter.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_name = "DiwakarTalwar/Anvi-0.01-test"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)

model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto"
)

model = PeftModel.from_pretrained(model, adapter_name)

The inference configuration may need to be adjusted for the available hardware and quantization setup.

Training Details

Training Data

The model was fine-tuned on a custom routing dataset containing natural-language requests paired with structured routing decisions.

The dataset covers programming, debugging, mathematics, weather, database queries, vision/OCR, translation, memory, task management, multi-step requests, and clarification.

Training Procedure

The model was fine-tuned using QLoRA, allowing a relatively small set of trainable adapter parameters while keeping the base model frozen.

The exported model uses PEFT-compatible LoRA weights.

Training Hyperparameters

Detailed hyperparameters are not documented in this release. Future versions may document learning rate, batch size, epochs, sequence length, optimizer, LoRA rank, LoRA alpha, dropout, and other training settings.

Training Hardware

Training hardware and total training time are not documented in this release.

Evaluation

Testing Data

Formal held-out evaluation results have not yet been published for this version.

Manual testing has been performed on routing prompts covering programming, debugging, mathematics, weather, SQL/database tasks, and structured-output instructions.

Planned Metrics

Future Anvi evaluations are intended to measure:

• —Task classification accuracy
• —Tool selection accuracy
• —Specialized-model selection accuracy
• —Valid JSON rate
• —Multi-step routing accuracy
• —Ambiguous-request handling
• —Hinglish request handling

No numerical benchmark results are claimed for this release.

Evaluation Status

Status: Experimental / preliminary

The current release demonstrates that QLoRA fine-tuning changes the model's instruction-following and routing behavior, but the routing protocol is not yet perfectly consistent.

A dedicated held-out routing benchmark is planned for future versions.

Technical Specifications

Model Architecture and Objective

Anvi is based on Qwen2.5-0.5B-Instruct and uses parameter-efficient LoRA fine-tuning.

The training objective is to map user requests to structured routing decisions rather than directly solve every request itself.

Software

• —Transformers
• —PEFT
• —Unsloth
• —PEFT version: 0.18.1

Citation

If you use Anvi in a project, reference the model repository:

DiwakarTalwar/Anvi-0.01-test

Glossary

QLoRA: A parameter-efficient fine-tuning method combining quantized base-model weights with trainable LoRA adapters.

LoRA: Low-Rank Adaptation, a method for fine-tuning a small set of adapter parameters instead of updating the complete base model.

Supervisor: A model whose primary role is deciding how a request should be handled by other models or tools.

Routing: Selecting the appropriate task, tool, model, or workflow for a user request.

Model Card Authors

Diwakar Talwar

Anvi is an independent experimental AI project.

Framework Versions

• —PEFT 0.18.1