metacogna/comPilar-Qwen3-4b-PilForce

Training Methodology

Phase 1: Pillar Grounding

The first fine-tuning phase focused on grounding the model in a stable vocabulary and conceptual space related to core expectation dimensions commonly observed in group coordination (e.g. progress, cooperation, inclusion, influence, reliability).

• Objective: establish consistent internal representations
• Emphasis: descriptive explanations, abstraction, and neutrality
• Format: single- and multi-turn instructional data
• Outcome: improved coherence and consistency in high-level explanations

---

Phase 2: Interaction and Contextual Reasoning

The second fine-tuning phase introduced interactional data that emphasizes how expectations influence one another across situations.

• Objective: enable relational reasoning rather than isolated explanations

• Emphasis: situational interpretation, contextual variation, and contrast

• Coverage:

  • expectation-to-expectation interactions
  • differences between egalitarian and hierarchical contexts
  • ambiguity and context-dependence

• Outcome: improved ability to explain how and why interpretations shift in real-world coordination scenarios

---

Parameter-Efficient Fine-Tuning (PEFT)

Training was performed using a parameter-efficient fine-tuning approach:

• Base model weights loaded in low-bit precision for memory efficiency
• Trainable adapters applied to attention projection layers
• Base weights frozen during training
• Gradients confined to adapter parameters

This approach preserves the base model’s general language capabilities while enabling targeted specialization in coordination dynamics.

---

Quantization

Training-Time Precision

• Base model loaded using low-bit quantization for efficiency
• Adapter training performed with higher-precision compute
• No permanent modification to base weights during training

Deployment Quantization

• After training, adapters were merged into the base model

• The merged model was quantized to 8-bit integer precision (Q8)

• Quantization selected to balance:

  • inference efficiency
  • retention of relational and interpretive nuance

The final artifact is a single, quantized model suitable for efficient local and edge deployment.

---

Intended Use

• Educational explanations of group and organizational dynamics
• Descriptive analysis of coordination patterns
• Exploratory discussion of social interpretation and context

Not Intended For

• Coaching, therapy, or managerial advice
• Prescriptive recommendations
• Moral or normative judgment
• Diagnostic or decision-making automation

---

Behavioral Characteristics

• Neutral, calm tone
• Context-sensitive depth (situational vs analytical)
• Explicit acknowledgment of uncertainty where appropriate
• Clear distinction between peer-based and role-based coordination contexts

---

Limitations

• Does not provide actionable guidance
• Explanations are interpretive, not predictive
• Outputs depend on clarity of contextual cues in the prompt

---

If you want, I can:

• Produce a registry-optimized short card
• Add a training diagram (textual)
• Write a deployment-focused variant for Ollama or similar runtimes