Add AIREV-Agent-0.8B v2: Sub-billion parameter model for BFCL V4

[mk42-ai]

Model

AIREV-Agent-0.8B — a 752M parameter model fine-tuned for agentic tool calling.

• Base: Qwen3.5-0.8B (Gated Delta Network architecture, 262K context)
• HuggingFace: https://huggingface.co/airev-ai/AIREV-Agent-0.8B
• License: Apache 2.0
• Organization: AIREV (On-Demand.io)

Training Pipeline

1. SFT on 50K Claude Opus 4.6-generated BFCL-format samples with chain-of-thought reasoning
2. AutoResearch — Karpathy-style automated hyperparameter discovery (112 experiments, 4 GPUs) found optimal GRPO config: lr=2e-6, 24 generations, temp=0.6, format_bonus=0.1
3. GRPO with AutoResearch-optimized config on 43K clean training samples (14 hours, single H100)
4. Targeted SFT on multi-turn, memory, and web_search categories using real BFCL function schemas

Evaluation

All 20 BFCL V4 categories evaluated. Results generated using transformers inference with temperature=0.6, chain-of-thought reasoning via tokens.

Prompt Mode

This model uses prompt-based function calling (not native FC mode). The BFCL system prompt with bracket format [func_name(params)] is used.

Hardware

Trained on a single NVIDIA H100 80GB GPU. Total training time: ~24 hours (SFT + GRPO + targeted SFT).