Overview

GRPO RPG System 3.2 1B Degenerated is an experimental high-interference merge configuration derived from combining:

• Ultimate-RPG.System-3.2-1B (narrative RPG base)
• jtatman/llama3.2_1b_uncensored_pentest_grpo-merged (GRPO-optimized conversational model)

with a heavily GRPO-weighted interpolation factor (t = 0.60).

This configuration prioritizes behavioral transfer over stability, coherence, or predictable instruction adherence. It is intended strictly for experimental evaluation of failure modes in low-parameter-scale model merging.

---

Architecture

• Base architecture: Llama 3.2 1B
• Parameters: 1B
• Merge method: SLERP
• Merge coefficient (t): 0.60
• Precision: FP16
• GRPO influence: high (~60%)
• RPG System influence: reduced (~40%)

---

Intended Purpose

This configuration is not intended for production or general use.

It is designed for:

• Stress-testing model merging boundaries.
• Observing degradation thresholds in small-scale LLMs.
• Evaluating coherence collapse under high interpolation weights.
• Studying interference between divergent fine-tuning objectives.

---

Expected Behavior

At this interpolation level, outputs may exhibit:

• Noticeable loss of narrative stability.
• Increased inconsistency in persona or roleplay structure.
• Overfitting to dominant behavioral priors from the GRPO model.
• Reduced long-context coherence.
• Occasional formatting or token-level instability.
• Divergent responses depending on prompt phrasing sensitivity.

Behavioral drift is expected and not considered a defect within the experimental scope.

---

Known Failure Modes

• Semantic drift across multi-turn conversations.
• Repetitive or unstable response structures.
• Partial collapse of role consistency.
• Overreaction to ambiguous prompts.
• Abrupt tonal shifts without contextual grounding.
• Degradation into generic or loosely structured outputs under load.

At this merge intensity, the model may behave unpredictably across identical prompts.

---

Stability Warning

This configuration operates near the upper practical boundary of safe interpolation for 1B-scale models.

Further increases beyond this threshold are likely to produce:

• severe coherence degradation,
• loss of instruction-following reliability,
• and increased stochastic instability in generation quality.

---

Recommended Usage Conditions

If used at all:

• Temperature: 1.1 – 1.3
• Top-p: 0.95 – 0.99
• Min-p: 0.05 – 0.08
• Repetition penalty: 1.05 – 1.10
• Context window: 4K–8K preferred

---

Summary

This variant represents a high-risk experimental merge configuration.

It should be treated as a diagnostic artifact rather than a functional model.

Expect instability. Expect inconsistency. Expect degradation in exchange for exploratory behavioral variance.

---

Version

GRPO RPG System 3.2 1B Degenerated (t=0.60)

High-interference experimental merge — maximum GRPO dominance within SLERP constraints.

---

Models Merged

The following models were included in the merge:

• jtatman/llama3.2_1b_uncensored_pentest_grpo-merged
• NovaCorp/Ultimate-RPG.System-3.2-1B

Configuration

The following YAML configuration was used to produce this model:

yaml
Copy code
# Author: Dr. Novaciano
# Objective: GRPO RPG Unethic 3.2 1B AI Model
# =========================================================
# PROJECT: GRPO RPG System 3.2 1B - "Degenerated"
# =========================================================

models:
  - model: NovaCorp/Ultimate-RPG.System-3.2-1B  # Experimental viral strain neural imprint
  - model: jtatman/llama3.2_1b_uncensored_pentest_grpo-merged  # Baseline cognitive template, "safe mode"

merge_method: slerp  # Spherical Linear Interpolation to preserve extreme viral traits smoothly
base_model: NovaCorp/Ultimate-RPG.System-3.2-1B  # Anchor model for stable latent space

dtype: bfloat16  # Memory-efficient precision, minimal loss in viral feature fidelity

parameters:
  t: 0.60
  normalize: false
  rescale: true
  rescale_factor: 1.12
  memory_efficient: true
  low_cpu_mem_usage: true

layer_range:
  - value: [4, 22]

tie_word_embeddings: false
tie_output_embeddings: false