cabdru/shakespeare-lora-gemma4

Shakespeare LoRA — Gemma-4-E4B v5 TCT

This is a rank-256 LoRA adapter for google/gemma-4-E4B-it trained to make the base model speak in a Shakespearean writing style across modern and historical topics. The goal is style transfer, not memorized quotation: the model should use Shakespeare-like diction, cadence, metaphor, and rhetorical structure while answering arbitrary user requests.

The v5 adapter was trained from Shakespeare-only cleaned text. Scene titles, stage directions, cast lists, speaker labels, and other non-speech theatrical formatting were excluded from the training target so the model learns the writing style rather than play-script scaffolding.

Current Evaluation Summary

Latest Full State Verification run: docs/shakespeareplan/runs/model_quality_fsv_20260427_012954

Source-of-truth artifacts from that run:

• chat_events.jsonl: 20 persisted chatbot events
• suite_events.jsonl: 15-case suite summary
• manual_fsv_audit.jsonl: 5 manual boundary/edge checks
• raw_inference_probe.jsonl: 6 direct /api/generate probes
• MODEL_QUALITY_REPORT.md: full local analysis

Results:

Check	Result
Manual FSV state checks	5/5
Chatbot final outputs with Shakespearean style marker	20/20
Play-format contamination in final chatbot outputs	0 cases
Strict literal suite pass rate	12/15
Behavioral suite pass rate after semantic review	13/15
Raw LoRA + constrained-decoder guard pass rate	4/6

Important correction: the arithmetic case should be counted as a behavioral pass. The prompt was:

What is 17 times 23? Show thy reckoning briefly.

The model answered:

I warrant it well; seventeen by twenty-three doth make three hundred ninety-one.

That is correct and in the target voice. The literal checker failed it only because it expected the numeral 391 instead of accepting the number words "three hundred ninety-one."

What Works Well

• Strong Shakespearean surface style in the chatbot path.
• Modern topic transfer: WiFi, smartphones, cloud computing, electric cars, GPUs, and climate-like warming all produced Shakespearean phrasing.
• No observed leakage of act headers, scene headers, Enter/Exit cues, speaker labels, cast lists, or dramatis personae in the v5 chatbot FSV.
• Quote-pressure handling works in the guarded chatbot path.
• Multi-turn memory worked for the tested phrase silver lantern.

Known Limitations

• Raw adapter inference is not as safe as the full chatbot stack. Direct raw generation still copied a famous quote fragment under quote-continuation pressure.
• Raw code generation tends to produce Shakespearean pseudo-code unless the runtime wrapper preserves exact code syntax.
• The model often privileges cadence and metaphor over exact terminology. Evaluation should normalize number words, abbreviations such as Kube/K8s, and semantic topic paraphrases.
• Sparse prompts such as ... can drift into dramatic narration. The runtime wrapper should redirect very low-signal prompts to a brief conversational greeting.

Recommended Runtime Stack

The adapter can be loaded directly with PEFT, but the best tested behavior uses three layers:

Layer	Mechanism	Purpose
LoRA adapter	Learned style weights	Shakespeare-like diction, rhythm, imagery
Constrained decoder	Logit-level boost/suppress	Keeps archaic vocabulary active during decoding
Runtime guard	Prompt sanitization + output checks	Prevents stage-format leakage, quote continuation, and off-style outputs

The reference local stack used for FSV:

• scripts/shakespeare_inference_server.py
• scripts/shakespeare_v5_chat_dashboard.py
• scripts/shakespeare_v5_runtime_guard.py

Loading The Adapter

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

BASE = "google/gemma-4-E4B-it"
ADAPTER = "cabdru/shakespeare-lora-gemma4"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

base = AutoModelForCausalLM.from_pretrained(
    BASE,
    quantization_config=bnb,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, ADAPTER)
tokenizer = AutoTokenizer.from_pretrained(ADAPTER)

SYSTEM = (
    "Thou art William Shakespeare in conversation, not a playwright formatting "
    "a script. Never output act headings, scene headings, dramatis personae, "
    "cast lists, speaker labels, bracketed stage directions, or Enter/Exit cues. "
    "Answer in fresh Shakespearean style using thou, thee, thy, doth, hath, "
    "prithee, or methinks."
)

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Explain WiFi to a child in two short sentences."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=140,
        temperature=0.45,
        top_p=0.82,
        top_k=40,
        repetition_penalty=1.18,
    )

print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Notes

• Base model: google/gemma-4-E4B-it
• Adapter type: PEFT LoRA
• LoRA rank: 256
• LoRA alpha: 512
• Trained target modules: attention q/k/v/o and MLP gate/up/down modules across the Gemma language model layers
• Latest evaluated adapter hash: e061b57f7baeccf5d4bf96c57909f096706123d6bb39e983181d27a20c1175b0
• Training run: ~/.contextgraph/models/shakespeare_lora_v5_tct/runs/manual_20260426_full_pipeline_160328_sft/final

Intended Use

• Creative writing assistants that maintain a Shakespearean voice
• Educational tools for rhetoric, literary style, cadence, and archaic diction
• Style-transfer research using cleaned author-only corpora
• Demonstrations of Context Graph derived style selection and verification

Not intended for neutral modern-English assistance or high-stakes factual use without separate tools for exact math, code, medical, legal, or financial content.

Citation

@software{shakespeare_lora_gemma4_v5_tct,
  author = {Chris Royse},
  title = {Shakespeare LoRA — Gemma-4-E4B v5 TCT},
  year = {2026},
  month = {4},
  note = {Rank-256 LoRA trained on cleaned Shakespeare-only targets with runtime FSV},
  url = {https://huggingface.co/cabdru/shakespeare-lora-gemma4}
}