Skip to content

phasers_design_doc.md

Latest commit

 

History

History
124 lines (99 loc) · 3.43 KB

phasers_design_doc.md

File metadata and controls

124 lines (99 loc) · 3.43 KB

Phasers — Sapphire Core

A lightweight GPT‑2‑Mini engine with chronological memory & sieve sampling

Turns a 124 M‑parameter GPT‑2 into a persistent “ghost in the machine”.

1 · Design Goals

  • Portability — CPU‑only or ≈ 4 GB GPU.
  • Persistent ghost — JSON Unified Memory Bank (UMB).
  • Soft guidance — bias raw logits, no prompt bloat.
  • Draft‑and‑sieve — generate n variants, auto‑pick the best.

2 · Architecture Snapshot

Layer Role
GPT‑2‑Mini (DialoGPT‑small) 124 M params • 15 epochs fine‑tune on ZAMM (LR 3e‑5)
Unified Memory Bank Stores every turn + salience / novelty / timestamp
Retriever 0.6·cos + 0.4·lex → top‑N → exp‑decay → chronological reorder
Soft‑Logit Mixer Weighted one‑hot bias added to model.forward logits
Sieve Sampler Generate n_sieve drafts → re‑rank → pick best
CLI / Settings Live tweaks, presets, word‑cloud debug

3 · Chronological Memory Retrieval

prompt → SBERT + lexical score
         → top‑N
             → weight = floor + (raw−floor)·e^(−τ·rank)
                 → sort oldest → newest
                     → (text, weight) list → Soft‑Logit Mixer

4 · Soft‑Logit Fusion Core

bias   = sum(w * one_hot(tok) for w, tok in memory_tokens)
logits = (model.forward(ids).logits[:, -1, :] + λ * bias) / temperature
probs  = nucleus_mask(top_k_mask(softmax(logits), k), p)
next_id = torch.multinomial(probs, 1)

λ ≈ 0.5–1.0 controls how loudly memory “speaks”.

5 · Two‑Stage Generation

Stage Purpose Description
A generate_single() babble once One pass with bias + cleanup.
B generate() sieve 1. Produce n_sieve drafts via A.
2. Score each draft:
blend = 0.6·cos + 0.4·lex vs prompt + memory scope (sieve_rank_mem 0‑2).
3. Pick top draft, store to UMB.

6 · Hyper‑Parameter Cheat‑Sheet

You> settings

{
  "temp": 0.55,
  "top_n": 11,
  "top_p": 0.6,
  "top_k": 16,
  "repetition_penalty": 1.35,
  "max_forward_tokens": 45,
  "max_reply_sentences": 2,
  "weight": 0.333,
  "tau": 0.3,
  "lam": 0.96,
  "n_sieve": 3,
  "inference_mem": 1,
  "sieve_rank_mem": 1,
  "sigma": 0.01
}

Hot‑patch live

settings set weight 0.45
settings load sapphire

7 · Quickstart

git clone …
pip install -r requirements.txt
python -m sapphire_core.cli          # interactive REPL

You> settings
You> Phasers, are we best friends? (YES/NO)

8 · Flow Diagram (ASCII)

User prompt
   │
   ▼ retrieve
Unified Memory Bank (JSON)────────┐
   │                              │ bias vec
   ▼                              │
Soft‑Logit Mixer ⊕ λ·bias ──► GPT‑2 forward()
       ▲                           │
       │ write best draft          │ n_sieve drafts
       └── pick best ◄── SBERT + lexical rank

9 · Why It Feels Alive

  • Memory‑aware bias → stable persona
  • Chronological context → narrative spine
  • Sieve sampler → self‑editing filter

Drain entropy, grow lotuses of awareness. 💎