# Agastya Decision RAG 2B by Andra Labs (Amartya Sai Andra)

> Agastya Decision RAG 2B offers open weights for typed decisions inside retrieval. Read its loading requirements and limits of the training record.

Updated 2026-11-07 (Australia/Adelaide)

## Quick answers

### What is Agastya Decision RAG 2B by Andra Labs (Amartya Sai Andra)?

[Agastya Decision RAG 2B](<https://decisionmodels.ai/models/agastya-rag-2b>): A Qwen3.5-2B fine-tune with an Unsloth Clef decision head, intended to choose sections in table-of-contents retrieval.

[Agastya Decision RAG 2B facts](<https://decisionmodels.ai/models/agastya-rag-2b>)

[Source](<https://huggingface.co/amartya16/Agastya-RAG-2B>)

### Who makes Agastya Decision RAG 2B?

[Agastya Decision RAG 2B](<https://decisionmodels.ai/models/agastya-rag-2b>): Andra Labs (Amartya Sai Andra)

[Agastya Decision RAG 2B facts](<https://decisionmodels.ai/models/agastya-rag-2b>)

[Source](<https://huggingface.co/amartya16/Agastya-RAG-2B>)

### Are Agastya Decision RAG 2B's weights open?

[Agastya Decision RAG 2B](<https://decisionmodels.ai/models/agastya-rag-2b>): Public weights; Unsloth FastDecisionModel required for typed inference; hosted handler deployment untested Licence: Apache-2.0

[Agastya Decision RAG 2B facts](<https://decisionmodels.ai/models/agastya-rag-2b>)

[Source](<https://huggingface.co/amartya16/Agastya-RAG-2B>)

By Andra Labs (Amartya Sai Andra) · Weights upload: 2026-11-06 · OPEN

A Qwen3.5-2B fine-tune with an Unsloth Clef decision head, intended to choose sections in table-of-contents retrieval. It was trained on general typed decisions. Retrieval, citations and final answer writing remain the application’s job; table-of-contents routing quality is unmeasured.

## Facts

- Base and size: Qwen3.5-2B; merged backbone plus separate decision-head weights
- Inputs: Typed state and choice, noul or score questions
- Context: Decision config records max\_len 4,096; backbone settings are not a tested task limit
- Licence: Apache-2.0
- Access: Public weights; Unsloth FastDecisionModel required for typed inference; hosted handler deployment untested

## Claims and sources

- **VENDOR-REPORTED**: The training record reports 66.0% accuracy and 0.0948 ECE over 150 held-out decisions, against 37.33% and 0.0967 for the base. The split, seed and evaluation script are absent. These are general decision results, not a retrieval benchmark. [source](<https://huggingface.co/amartya16/Agastya-RAG-2B>)

## Links

- [Source: Model card and weights](<https://huggingface.co/amartya16/Agastya-RAG-2B>)

## Related projects

No related projects are listed in the registry.