Eyght/eyght-p1

Eyght P1 — Phone AGI Model

The first phone-native AGI model trained with 315 GRPO reward functions.

Built by Eyght · Powered by Veta Art Training Studio · RTX 4070 Ti

---

📱 Overview

Eyght P1 is a 1.5B parameter language model fine-tuned using Group Relative Policy Optimization (GRPO) with 315 reward functions across 85 datasets. It is designed to run 100% on-device on mobile phones via PocketPal AI, with zero cloud dependency.

Property	Value
Base Model	Qwen2.5-1.5B-Instruct
Parameters	1.58B (36.9M trainable via LoRA r=32)
Quantization	Q4_K_M (GGUF)
File Size	940 MB
Context Length	32,768 tokens
Architecture	qwen2 (decoder-only transformer)
Training Method	GRPO (Group Relative Policy Optimization)
Training Steps	500
Generations per Prompt	8
Max Completion Length	2,000 tokens
Hardware	NVIDIA RTX 4070 Ti (12GB)
Training Time	~2 hours
License	Apache 2.0

---

🧠 Capabilities

Eyght P1 is a full AGI model — not just a coder. It was trained across all major AI domains:

Domain	Datasets	Description
🧠 AGI Reasoning	19	Chain-of-thought, deep thinking, ARC-AGI grid reasoning, AGIEval exam reasoning, spatial reasoning, RAG
💻 Coding	8	Code debugging, HumanEval, SWE-bench, Codeforces, open-code reasoning, IDE tools
🤖 Agent/Tool	27	Terminal, tool calling, agent trajectories, MCP, dynamic workflows, multi-step execution
💬 General	9	Conversation, creative writing, social/emotional intelligence, common sense
👁 Vision	5	Spatial reasoning, OCR, real UI understanding, vision-code
🔒 Safety	8	Security, red team, safety alignment, cyber patch
🌍 Multilingual	5	Translation, multilingual expanded
📚 Knowledge	17	MMLU, math, Wikipedia, SciQ, BBH, OpenBookQA, IFEval
Total	85	Full domain coverage

---

⚡ GRPO Training Details

Reward System

• 315 reward functions consolidated from 2,788 micro-rewards
• 16 high-signal reward functions active during training (balanced_15 preset + anti-cheat)
• Reward categories: code quality, reasoning, math accuracy, safety, efficiency, self-correction, knowledge, anti-hacking, attention, web apps, process supervision, format compliance

Key GRPO Techniques Applied

Technique	Reward ID	Description
Temperature-Diverse Group Sampling	t265	Heterogeneous temperature schedule prevents homogeneous groups
Dual-Anchor Advantages	t266	Absolute goal-aware anchor + group-relative advantage
Prospect-Theoretic Advantage Shaping	t267	Bounded asymmetric shaping from Prospect Theory
Homogeneous Group Collapse Prevention	t268	Prevents zero-variance groups in GRPO
Multiplicative Reward Decomposition	t241	Multiply (not add) reward components
Dynamic Sampling Zero-Variance Filter	t244	DAPO filtering of zero-variance groups
Clip-Higher Asymmetric Clipping	t245	Asymmetric clipping for exploration
Token-Level PG Loss	t246	Better for long reasoning traces
Anti-Cheat	t236	Always-on reward gaming prevention

Training Configuration

base_model: unsloth/Qwen2.5-1.5B-Instruct
lora_r: 32
lora_alpha: 32
lora_dropout: 0.05
learning_rate: 5e-6
max_steps: 500
num_generations: 8
batch_size: 8
gradient_accumulation: 4
max_prompt_length: 384
max_completion_length: 2000
max_seq_length: 2560
temperature: 0.9
beta: 0.0
loss: dapo
gradient_checkpointing: unsloth

---

📊 Benchmark Coverage

This model was trained on data covering the following benchmarks:

Benchmark	Type	In Training
ARC-AGI-2	Abstract grid reasoning	✅ (165 tasks + 5,000 programs)
AGIEval	Exam reasoning (SAT, Gaokao, LSAT)	✅ (2,331 examples)
GSM8K	Math word problems	✅
HumanEval	Code generation	✅
SWE-bench	Software engineering	✅
MMLU	General knowledge	✅
BBH	Big-bench hard	✅
IFEval	Instruction following	✅
Codeforces	Competitive programming	✅

---

📱 Deployment

PocketPal AI (Recommended)

1. Install PocketPal AI on your phone (iOS / Android)
2. Open the app → Models → Search HuggingFace
3. Search for Eyght/eyght-p1
4. Download → Load → Start chatting!
5. 100% private, no cloud, no internet needed after download

Minimum Phone Requirements

Requirement	Minimum	Recommended
RAM	3 GB free	4 GB+ free
Storage	1 GB free	2 GB free
CPU	Any	Apple Silicon (Metal) / Qualcomm (NPU)

Ollama (Desktop)

ollama run eyght-p1

---

🔧 Technical Details

Training Pipeline

1. Data Loading: 85 JSONL datasets loaded → 885,945 total examples
2. Reward Selection: 315 rewards → 16 high-signal functions consolidated
3. GRPO Training: 500 steps, 8 generations per prompt, DAPO loss
4. LoRA Adapter: r=32, alpha=32, all attention + MLP modules
5. Merge: LoRA adapter merged with base model (full precision)
6. Quantization: Merged model → Q4_K_M GGUF (940 MB)
7. Deployment: GGUF → Ollama + HuggingFace upload

Reward Catalog

This model uses the Eyght Reward Catalog — a comprehensive collection of 315 GRPO reward functions covering:

• Chinese models: Qwen3.8-Max, AgentWorld, DeepSeek, MiniMax, Kimi, GLM, MAI
• US models: GPT-6, Gemini, Claude, Grok, Nemotron, Skywork
• Abu Dhabi models: MBZUAI
• Research papers: MDP-GRPO, DAPO, Search-R1, ToolRL, KTO, DPO, SRPO, ARiSE
• Custom rewards: ARC-AGI reasoning, AGIEval, temporal fact-currency, MLOps agent

Presets Available

Preset	Rewards	Focus
balanced_15	17	General purpose (used for this model)
code_focused	17	Software engineering
reasoning_focused	14	Math + logic
mdp_grpo	16	MDP-GRPO 3 components
qwen_frontier	24	Qwen3.8 capabilities
grpo_techniques	8	GRPO training techniques
rl_developments	38	All RL developments
agi_reasoning	9	ARC + AGIEval
financial_grpo	5	Credit underwriting
math_reasoning_tools	13	Math + tools
grpo_models	15	GRPO model techniques

---

⚠️ Limitations

• Model size: 1.5B parameters — less capable than 7B+ models on complex tasks
• Phone hardware: Performance depends on phone RAM and CPU/NPU
• Training data: Trained on available open datasets; may not cover all domains equally
• Reward design: GRPO rewards are heuristic-based; may not capture all quality dimensions
• Not for production: This is a research/personal-use model, not a commercial product

---

🏗️ Built With

Tool	Role
Unsloth	2x faster fine-tuning
TRL	GRPO trainer
Veta Art Studio	Training orchestration
PocketPal AI	Phone deployment
Ollama	Desktop deployment
llama.cpp	GGUF inference

---

📄 License

Apache 2.0 — See LICENSE for details.

---

🙏 Acknowledgments

• Qwen team (Alibaba) for the excellent Qwen2.5-1.5B base model
• Unsloth team for making fine-tuning 2x faster and accessible
• PocketPal AI (Asghar Ghorbani) for the mobile inference platform
• Hugging Face for model hosting and the TRL library
• The open-source AI community for the datasets and research

---

Eyght P1 — Your AGI in your pocket. Built by AI, for you. 📱🤖