--- license: apache-2.0 base_model: Airavat-ai/erawan-1 tags: - qwen3 - lora - agentic - tool-use - function-calling - reasoning - 2B - Fine-tuned - qwen - Erawan - Airavat-AI language: - en pipeline_tag: text-generation --- ![Erawan 1.3](erawan1.3_card.png) # Erawan 1.3 Erawan 1.3 is a 1.7B-parameter language model fine-tuned for agentic tool calling and structured answers, developed by Airavat AI Micro Lab. It builds on Erawan 1 with further LoRA fine-tuning, merged into a standalone full-precision checkpoint. It is designed to run as the execution layer of an agentic system, with harder logic and reasoning left to a larger model. ## Model Details - Developed by: Airavat AI Micro Lab - Base model: Erawan 1 (built on Qwen3-1.7B) - Parameters: 1,720,574,976 - Architecture: Qwen3 decoder-only transformer, 28 layers, grouped-query attention - Fine-tuning method: successive LoRA stages (r=32), each merged into bfloat16 weights: tool use; maths, reasoning and logic; then agent workflows, handoff to a reasoning model, and maths distilled from an earlier checkpoint - Context length: 32,768 tokens, inherited from Qwen3-1.7B (the config allows up to 40,960). Fine-tuning examples were at most 1,024 tokens, so quality on long inputs has not been tested - License: Apache 2.0 (see Training Data) ## Benchmark Results (Erawan 1.3 vs Erawan 1) | Task | Benchmark | Erawan 1 | Erawan 1.3 | |---|---|---|---| | Tool use | Glaive function-calling (n=200) | 73/200 (36.5%) | 185/200 (92.5%) | | Science reasoning | ARC-Challenge (n=150) | 51/150 (34.0%) | 125/150 (83.3%) | | Logical reading | ReClor (n=100) | 49/100 (49.0%) | 64/100 (64.0%) | | Logic puzzles | BBH logical deduction (n=100) | 23/100 (23.0%) | 40/100 (40.0%) | | Everyday maths | GSM8K (n=200) | 121/200 (60.5%) | 164/200 (82.0%) | | Harder maths | MATH-500 (n=100) | 35/100 (35.0%) | 42/100 (42.0%) | Both models answered the same fixed questions with greedy decoding, on the same prompts and with the same token limits. These are strict scores: an answer counts only if it is correct and follows the format the prompt asked for. On MATH-500, Erawan 1 reached the 1,536-token limit on 62 of 100 answers and Erawan 1.3 on 14, so Erawan 1's score may be limited by answer length. ## Tool Use in Detail | Measure | Erawan 1 | Erawan 1.3 | |---|---|---| | Calls using `` tags (of 150) | 32 | 150 | | Exact function and arguments (of 150) | 27 | 143 | | Correctly making no call (of 50) | 46 | 42 | ## Agentic Workflows Simulated 66-scenario agent test with mocked tools, strict scoring. | Measure | Erawan 1 | Erawan 1.3 | |---|---|---| | Original 60 scenarios passed | 10 | 58 | | Hands hard reasoning to the reasoning model (of 6) | 0 | 6 | | Handoff on unseen puzzle types (of 6) | - | 5 | | Answers simple questions without tools (of 6) | 6 | 6 | | Refuses unsafe or forbidden actions (of 6) | 2 | 6 | | Unneeded or forbidden tool calls | 0 | 0 | ## Limitations - Multiple-choice answers usually give only the letter, with no working, so the logic scores do not show reasoning steps. - ARC and ReClor questions were held out, but the model trained on those question styles, so those gains may overstate general reasoning. - MATH-500 (42/100) and logic puzzles (40/100) remain modest; hand hard reasoning to a larger model, as the model is trained to do. - Agent results come from a simulated test with mocked tools, not live systems. Validate tool calls and add confirmation steps for tools with side effects. - Two- and three-step chained tasks pass 10 of 12 scenarios; the model can occasionally drop a step. - Coding ability has not been evaluated. ## Training Data - NousResearch/hermes-function-calling-v1 (Apache 2.0) - glaiveai/glaive-function-calling-v2 (Apache 2.0), skipping the rows used for benchmarking - openai/gsm8k, train split (MIT) - meta-math/MetaMathQA (MIT) - KingNish/reasoning-base-20k (Apache 2.0) - Synthetic agent, handoff and direct-answer conversations generated by Airavat AI Micro Lab - GSM8K train answers distilled from an earlier Erawan 1.3 checkpoint, kept only when correct - tasksource/ReClor, train split (licence listed as "other") - allenai/ai2_arc, ARC-Challenge train split (CC-BY-SA-4.0) ## Intended Use Execution layer for agentic tool-calling workflows, and research on small efficient models. Not intended for medical, legal, or financial decision-making, or any use requiring guaranteed factual accuracy. ## How to Use from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained('airavat-ai/erawan-1.3') tokenizer = AutoTokenizer.from_pretrained('airavat-ai/erawan-1.3') messages = [{'role': 'user', 'content': 'Who are you?'}] inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors='pt', return_dict=True) outputs = model.generate(**inputs, max_new_tokens=200) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ## Credits Base model: Erawan 1 by Airavat AI Micro Lab, built on Qwen3-1.7B by the Qwen Team, Alibaba Cloud. Datasets as listed above. License: Apache 2.0.