DragonBophades/Inkling-Gutenberg-DPO

Inkling-Gutenberg-DPO

A merged BF16 release of Thinking Machines Inkling, preference-tuned to favor authentic literary prose over synthetic creative-writing slop.

This is the deployment-ready counterpart to `nbeerbower/Inkling-Gutenberg-DPO-LoRA`. The rank-32 adapter has already been merged into the complete Inkling checkpoint; no PEFT runtime or separate base-model download is required.

Training

Training used Direct Preference Optimization through Tinker and tinker-cookbook.

The selected pairs contrast public-domain Project Gutenberg prose (chosen) with synthetic prose (rejected). Alembic's deterministic quality score was used for selection; its raw LLM-judge preference vote was not used.

Training result

The preference signal converged rapidly, so the run was limited to one decaying-learning-rate epoch.

Format and use

This repository contains the full merged BF16 checkpoint in Inkling's original TML-native sharded layout. It is approximately 1.8 TiB and includes the unmodified MTP tensors from the base model.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "nbeerbower/Inkling-Gutenberg-DPO"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

Inkling is a 975B-total / 41B-active mixture-of-experts model. Loading or serving the BF16 checkpoint requires unusually large storage and memory resources. Use the LoRA repository if distributing only the tuned delta is preferable.

Merge validation

The merge applied the inverse of Transformers' Inkling checkpoint conversion to preserve the published native layout, including dense, routed-expert, and shared-expert fused/interleaved MLP tensors.

• —1,438 adapter tensors paired into 719 LoRA modules
• —591 native base tensors updated across all 108 model shards
• —all 1,552 indexed checkpoint tensors and 109 safetensors files header-validated after merge
• —representative attention and dense-MLP tensors matched the expected BF16 LoRA result exactly
• —LoRA scale: alpha / rank = 1

Intended behavior

The tune targets literary fiction, period prose, novel continuation, dialogue, and creative-writing tasks where generic AI phrasing is undesirable. It is a stylistic preference tune, not a factual-knowledge or safety tune.

The model may reproduce public-domain literary styles, favor longer source-like continuations, or use period diction when prompted toward historical settings. Evaluate modern-register writing and general instruction following for your application.

Recommended inference settings

The +21.03 reward margin above reflects a strongly peaked policy — the model commits hard to its preferred register. In practice this means it wants sampling: greedy decoding falls into repetition loops. The settings below were validated on the quantized int4 conversion through an OpenAI-compatible server, and apply to the BF16 checkpoint as ordinary sampling guidance.

• —Sample through the chat template; raw text completion is off-distribution for a chat-format DPO and degenerates.
• —temperature 0.7 is the most controlled and coherent; 0.9 is richer with occasional long sentences; >= 1.0 becomes vivid but breathless.
• —Avoid greedy / temperature 0.

Example at temperature 0.7, top_p 0.9, repetition_penalty 1.2 — "Write the opening paragraph of a novel about an old lighthouse keeper on his final night of duty.":

The lantern had burned for forty-seven years without fail, and Silas knew—without checking the clock or consulting any log—that tonight it would not be lit again by his hand. He climbed to its crown with legs that remembered every rung more clearly than they did stairs in town, carrying only one match he never intended to strike.

License

Apache 2.0, matching the Inkling base model. Alembic-DPO is CC-BY-4.0 and derives its literary text from public-domain Project Gutenberg sources; consult the dataset card for complete provenance.