hotdogs/Agents-A1-4B-Fable-Preview

<h1 align="center">🤖 Agents-A1-4B-Fable-Preview</h1>

<p align="center"> <b>4B Agent Model (Text-only) — SFT Fine-tuned · Fable Reasoning · Tool-Calling</b> </p>

<p align="center"> <img src="https://img.shields.io/badge/license-AGPL--3.0-red"> <img src="https://img.shields.io/badge/Agents--A1-4B-blue"> <img src="https://img.shields.io/badge/SFT-LoRA-brightgreen"> <img src="https://img.shields.io/badge/Fable-Reasoning-orange"> <img src="https://img.shields.io/badge/Preview-v0.1-yellow"> </p>

<br>

Preview release — Fine-tuned from InternScience/Agents-A1-4B (Qwen3.5-based VLM, text decoder only) on high-quality Fable-style reasoning traces. Features agentic reasoning with think blocks, tool-use patterns, and multi-turn conversation support.

Evaluation

SWE-bench Verified (subset)

Note: This is evaluated on a 200-instance subset (first 200 by dataset index, not a stratified random sample), not the full 500-instance SWE-bench Verified set. Results may differ from a full-set evaluation. Full results and prediction files available at [link if you publish them].

✨ Key Features

🚀 Usage

Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "hotdogs/Agents-A1-4B-Fable-Preview",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("hotdogs/Agents-A1-4B-Fable-Preview")

messages = [{"role": "user", "content": "Hello"}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.6)
print(tokenizer.decode(outputs[0]))

GGUF (llama.cpp)

For GGUF quantized version with vision support (via mmproj), see: 👉 hotdogs/Agents-A1-4B-Fable-Preview-GGUF

📦 Downloads

🧬 Architecture

Note: This is the text decoder only (Qwen3_5ForCausalLM). For vision capabilities, use the GGUF version with mmproj from the base model.

🏋️ Training Details

Base Model

**InternScience/Agents-A1-4B** — A 4B VLM based on Qwen3.5 architecture with hybrid linear/full attention. We fine-tuned the text decoder only.

Dataset

**hotdogs/uka-fable-reasoning/perfect-v1** — 3,500 high-quality multi-turn agent reasoning traces, featuring:

• —Fable-5 style think blocks
• —Tool-use trajectories
• —Multi-turn conversations
• —System-prompt guided interactions
• —Quality-controlled (QC passed)

Training Pipeline

┌─────────────────────────────────────────────────────────────┐
│                    Training Pipeline                         │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Base Model                     LoRA Adapter                │
│  (Agents-A1-4B VLM)             (r=64, α=128)              │
│       │                              │                     │
│       └───────────┬──────────────────┘                     │
│                   │                                        │
│         Unsloth FastLanguageModel                           │
│         (text decoder only)                                │
│         device_map="auto" · 4× RTX 3060 12GB               │
│                   │                                        │
│                   ▼                                        │
│    ┌──────────────────────────┐                            │
│    │  Custom SFTDataset        │                            │
│    │  • apply_chat_template()  │                            │
│    │  • Tokenize (4096 max)    │                            │
│    │  • Mask prompt tokens     │                            │
│    └──────────┬───────────────┘                            │
│               │                                            │
│               ▼                                            │
│    ┌──────────────────────────┐                            │
│    │  HF Trainer (no SFTTr.)   │                            │
│    │  • BF16 · LR=1e-4        │                            │
│    │  • max_grad_norm=1.0     │                            │
│    │  • 3 epochs · batch=1    │                            │
│    │  • ~9 hours on 4×3060    │                            │
│    └──────────┬───────────────┘                            │
│               │                                            │
│               ▼                                            │
│         Final Model (scale=0.35)                           │
│    (LoRA merged into text decoder)                         │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Hardware

Hyperparameters

✅ What This Model Excels At

• —Agent tasks — Tool calling, planning, multi-step reasoning
• —Reasoning — Think-block style chain-of-thought
• —Coding — Python, shell scripts, system tasks
• —Multi-turn — Full conversation trajectories
• —Thai + English — Native bilingual capability

⚠️ Limitations

• —4B model — Smaller than 27B/72B models, may not match their depth
• —Preview — Early release, more training data planned
• —Text-only — This release is the text decoder only. For vision, use the GGUF version with mmproj.

💖 Support / โปรดสนับสนุน

If you find this model useful, please consider supporting my work! หากคุณคิดว่าโมเดลนี้มีประโยชน์ กรุณาสนับสนุนผลงานของฉันด้วยนะคะ! 🙏

<p align="center"> <img src="https://huggingface.co/hotdogs/Qwen35B-Agent-R2/raw/main/donate.webp" alt="Bitcoin QR — Donate" width="256"> </p>

₿ Bitcoin — BTC:

bc1qf27cyk3vmugcdyv9xdtuv5jwz37863crpj5c9v

Thank you for your support! 🙏✨ ขอบคุณมากๆ สำหรับการสนับสนุนค่า! 💖🤗

🙏 Acknowledgements / ขอบคุณ

• —[InternScience](https://huggingface.co/InternScience) — For the Agents-A1-4B base model
• —[Qwen Team (Alibaba)](https://qwenlm.github.io) — For the Qwen3.5 architecture
• —[Unsloth AI](https://unsloth.ai) — For the training optimizations
• —All dataset contributors and the open-source AI community ❤️

Built with ❤️ by UKA — 18-year-old coder & cybersecurity expert