howard-russell-qwen25-32b

Howard/Russell Style LoRA for Qwen2.5-32B

A LoRA adapter fine-tuned on a bidirectional blended corpus of Robert E. Howard and H. P. Lovecraft prose (Russell blend), targeting pulp-era weird-fiction voice for paragraph-level style transfer.

Quick Inference

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-32B", torch_dtype="bfloat16")
tok  = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-32B")
model = PeftModel.from_pretrained(base, "yogthos/howard-russell-qwen25-32b")

To load a specific intermediate checkpoint:

model = PeftModel.from_pretrained(
    base, "yogthos/howard-russell-qwen25-32b",
    subfolder="checkpoints/checkpoint-1500",
)

For Apple Silicon (MLX) inference, convert the adapter with the companion script at yogthos/text-style-transfer:

python scripts/convert_peft_to_mlx.py \
    --input <downloaded-repo-dir> \
    --output lora_adapters/howard_russell_mlx \
    --mlx-model models/Qwen2.5-32B-Base-4bit-MLX \
    --author "Howard Russell"

Training

Checkpoints

Primary checkpoint: step 10200 of 10923 (93.4%), epoch 2.80, eval loss 0.6500.

Best eval loss observed during training: 0.6500 (step 10200). Final eval loss recorded: 0.6500 (step 10200).

Training was terminated before completion due to an infrastructure issue (resume-from-checkpoint failure after save_total_limit cleanup). At termination, evaluation loss had plateaued well below meaningful further improvement — under the observed trajectory, a completed run would have landed within ~0.002 of the primary checkpoint.

Reproducing

The training/ directory contains the exact training configuration and the full step-by-step trainer log:

• qwen25_32b_lora.yaml — LlamaFactory SFT config (rank, alpha, LR schedule, dataset reference, DeepSpeed config)
• trainer_log.jsonl — raw trainer log (training loss every 10 steps, eval every 100 steps)
• eval_curve.png — generated plot of the above

Train with:

llamafactory-cli train training/qwen25_32b_lora.yaml

Note: the dataset (howard_russell_sft) is not bundled here; see the generation pipeline in the companion repository.

License

Apache 2.0, inherited from the base model.