---
base_model: thinkingmachines/Inkling
library_name: peft
license: apache-2.0
language:
- en
pipeline_tag: text-generation
datasets:
- schneewolflabs/Alembic-DPO
tags:
- tinker
- tinker-cookbook
- peft
- lora
- dpo
- inkling
- gutenberg
- creative-writing
- anti-slop
---

# Inkling-Gutenberg-DPO-LoRA

A rank-32 LoRA for [Thinking Machines Inkling](https://huggingface.co/thinkingmachines/Inkling), trained to prefer authentic literary prose over synthetic creative-writing slop.

The adapter is the preference-tuned component of the nbeerbower Gutenberg series. It preserves Inkling as the general-purpose base while applying a focused literary-fiction bias: stronger narrative texture and interiority, more controlled pacing, and an active dispreference for formulaic AI phrasing.

## Training

Training used Direct Preference Optimization through [Tinker](https://thinkingmachines.ai/tinker) and `tinker-cookbook`.

| Setting | Value |
|---|---:|
| Base model | `thinkingmachines/Inkling` |
| Dataset | `schneewolflabs/Alembic-DPO`, `scored` configuration |
| Selection | English Gutenberg, `keep`, `quality >= 50` |
| Train / validation pairs | 3,719 / 128 |
| Objective | DPO |
| LoRA rank / alpha | 32 / 32 |
| DPO beta | 0.1 |
| Learning rate | `1e-5`, linear decay |
| Effective pair batch | 32 |
| Maximum sequence length | 4,096 |
| Epochs | 1 (116 optimizer steps) |
| Renderer | Inkling `tml_v0`, thinking effort 0.9 |

The selected data pairs public-domain Gutenberg prose (`chosen`) with synthetic prose (`rejected`). Alembic's deterministic quality scoring was used for selection; its raw LLM-judge preference vote was not used because reconstructed Gutenberg prompts can reward prompt adherence over fidelity to the source literature.

## Training result

| Metric | Start | Final |
|---|---:|---:|
| Held-out NLL | 1.7854 | 1.6891 |
| Preference accuracy | ~50% | 100% |
| DPO loss | ~1.8 | 0.0050 |
| Reward margin | near 0 | +21.03 |

The preference signal converged rapidly, as expected for a capable base model and a high-contrast dataset. The single decaying-learning-rate epoch was retained; no additional epochs were run.

## Format and use

This repository contains a standard PEFT LoRA adapter. Load it together with the unmodified `thinkingmachines/Inkling` base model using a PEFT-compatible runtime.

```python
from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "thinkingmachines/Inkling",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, "nbeerbower/Inkling-Gutenberg-DPO-LoRA")
```

Inkling is a 975B-total / 41B-active MoE. This adapter includes expert-layer LoRA tensors and is consequently large. MoE expert LoRA serving remains experimental in some runtimes; a merged model is the most portable deployment format.

## Intended behavior

The tune is intended for literary fiction, period prose, novel continuation, dialogue, and creative-writing tasks where generic AI phrasing is undesirable. It is a stylistic preference adapter, 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.

## 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.