la-writer is a LoRA fine-tune of stabilityai/stablelm-2-1_6b-chat, merged into the base weights, trained to write literary analysis in a specific voice: short punchy openings, rhetorical questions as pivots, sparing use of ellipses and em dashes, dry understated reactions, no inflated claims of significance.

Model Details

Base model: stabilityai/stablelm-2-1_6b-chat (~1.6B params)
Fine-tuning method: LoRA (PEFT), merged into base weights for standalone use
Language: English
License: Inherits the base model's license (stabilityai/stablelm-2-1_6b-chat, custom "other" license) — see the base model license for terms before commercial use.

Model Description

This model was fine-tuned to produce literary-analysis prose in a specific, terse, first-person critical voice, guided by the system prompt:

Write literary analysis in this voice: open with a short punchy line, no throat-clearing. Use rhetorical questions to pivot. Allow ellipses for a beat and em dashes for asides. Drop short dry reactions. Do not inflate significance. Be specific and direct.

The repository previously hosted only the LoRA adapter weights; this update merges the adapter into the base model so it can be loaded directly with AutoModelForCausalLM, without requiring peft at inference time.

Uses

Direct Use

Generating literary analysis / close-reading commentary in the trained voice, e.g. as a writing assistant for essays, reviews, or annotations on literary texts. Best used with the system prompt above (see chat_template.jinja) for consistent style.

Out-of-Scope Use

Not intended as a general-purpose chat assistant, factual Q&A system, or for tasks requiring reliable factual accuracy — the base model's general instruction-following ability may be degraded by this narrow fine-tune.
Not evaluated for safety-critical, medical, legal, or financial use.
Subject to the base model's license restrictions on commercial use — check before deploying in a commercial product.

Bias, Risks, and Limitations

Very small training set. The model was fine-tuned on 23 examples for 5 epochs, which is enough to imprint a stylistic voice but is a high-overfitting regime — expect the model to closely echo phrasing patterns from the training examples rather than generalize broadly.
Inherits any biases and limitations of the base stablelm-2-1_6b-chat model.
Not evaluated on any held-out benchmark; only training loss was tracked.

How to Get Started with the Model

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Xx-Vexento-xX/la-writer")
tokenizer = AutoTokenizer.from_pretrained("Xx-Vexento-xX/la-writer")

messages = [
    {"role": "system", "content": "Write literary analysis in this voice: open with a short punchy line, no throat-clearing. Use rhetorical questions to pivot. Allow ellipses for a beat and em dashes for asides. Drop short dry reactions. Do not inflate significance. Be specific and direct."},
    {"role": "user", "content": "Analyze the ending of The Great Gatsby."},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
output = model.generate(inputs, max_new_tokens=300)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Training Data

23 hand-curated literary-analysis text examples (JSONL, single text field per example), formatted with the system voice prompt above.

Training Procedure

Method: LoRA via PEFT, r=4, lora_alpha=16, lora_dropout=0.05, target modules q_proj, v_proj, task_type=CAUSAL_LM
Trainable parameters: 786,432 (0.0478% of 1,645,301,760 total)
Epochs: 5
Batch size: 1 (per device) x 8 gradient accumulation steps = effective batch size 8
Learning rate: 2e-4, cosine schedule, 10% warmup
Precision: fp32
Hardware: CPU
Training regime: fp32, CPU-only

Speeds, Sizes, Times

Train runtime: ~954s (~16 min) for 15 optimizer steps
Final train loss: 3.359 (loss trended from ~3.64 at epoch ~0.35 down to ~3.00 at epoch 5)

Framework versions

PEFT 0.19.1
Transformers (see base model card for compatible versions)

Model Card Contact

Xx-Vexento-xX (via Hugging Face)