YoussefElsafi/PlayerAI-1.2B-v1.5

PlayerAI-1.2B-v1.5 is an enhanced version of PlayerAI, a fine-tuned conversational language model designed for immersive, human-like interaction in multiplayer social environments.

This release significantly improves conversational coherence, tone stability, and multi-turn context handling compared to earlier versions, with training extended to ~11k steps on a large-scale conversational dataset.

---

Model Overview

• Base Model: LiquidAI/LFM2.5-1.2B-Instruct
• Parameters: ~1.2B
• Architecture: Decoder-only Transformer
• Training Type: Supervised fine-tuning (full model)
• Training Scale: ~3–4M conversational pairs, ~11k steps
• Context Style: Multi-turn conversational sequences
• Primary Objective: Social realism in dialogue generation

---

What's New in v1.5

• Improved multi-turn coherence and memory retention
• More stable conversational tone across longer exchanges
• Reduced randomness and better response consistency
• Stronger adherence to chat formatting (User: / AI:)
• Overall more natural and believable dialogue flow

This update focuses on refinement and stability, not changing the core behavior of the model.

---

Intended Use

This model is intended for research and experimental use cases involving:

• Multiplayer conversational agents
• Social simulation environments
• NPC dialogue systems
• Human-like chat behavior modeling
• Interactive roleplay systems

It is not intended for:

• factual question answering
• structured instruction following
• safety-critical systems
• deterministic reasoning tasks

---

Training Data

The model was trained on a large-scale collection of synthetically generated and curated conversational sequences designed to replicate natural human-like chat behavior in multiplayer-style environments.

The dataset emphasizes:

• informal conversation structure
• rapid topic switching
• multi-turn dialogue continuity
• noisy and unstructured chat patterns
• social interaction realism over factual accuracy

No personally identifiable or sensitive user-specific content is included.

---

Chat Format

Training samples use a newline-based structured format (\n) representing conversation turns.

Format Structure

input: <conversation history>\nAI:
output: <next assistant response>

Conversation turns inside the input:

AI: <message>\nUser: <message>\nAI: <message>\nUser: <message>

The model predicts the next assistant response given the full conversation history.

---

Example Interactions

Note: All assistant messages are generated by PlayerAI-1.2B-v1.5.

Example 1 — Single Turn

Example 2 — Short Conversation

Example 3 — Extended Context Chain

Example 4 — Nonsense Interaction

Example 5 — Reverse psychology

---

Training Objective

The model is optimized to:

• maintain coherence across multi-turn dialogue
• generate short, informal responses
• adapt dynamically to conversational tone
• handle noisy and inconsistent chat structures
• simulate realistic multiplayer chat behavior

Loss is applied only on assistant outputs.

---

Inference

Basic Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "YoussefElsafi/PlayerAI-1.2B-v1.5"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

tokenizer.pad_token = tokenizer.eos_token

input_text = "User: wsp\nAI:"

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=80,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
        use_cache=True
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Multi-Turn Example

input_text = (
    "User: hi\\n"
    "AI: hello\\n"
    "User: how are you\\n"
    "AI: good u?\\n"
    "User: also good, what is 2+2?\n"
    "AI:"
)

---

Behavior Characteristics

• informal conversational tone
• short and adaptive responses
• occasional ambiguity or inconsistency
• strong dependence on recent dialogue context
• dynamic emotional and linguistic variation

These behaviors are intentional.

---

Limitations

• Not suitable for factual reasoning
• May become inconsistent in very long contexts
• Not optimized for structured instruction following
• Can exhibit conversational drift

---

Ethical Considerations

This model is designed for simulation and research.

• Outputs may appear human-like
• Behavior prioritizes realism over correctness
• Ambiguity is an intentional feature

Use appropriate safeguards depending on your application.

---

Attribution (Optional)

If you use PlayerAI in a project, attribution is appreciated but not required:

"Powered by PlayerAI"

---

License

Apache 2.0