A LoRA fine-tune of Qwen3-1.7B specialized on bishop concepts and strategy in chess (annotated games, study prose, and Q&A). This repo ships both the PEFT LoRA adapter and a quantized GGUF you can run directly in Ollama / llama.cpp.

Base model: unsloth/qwen3-1.7b-unsloth-bnb-4bit (Qwen3-1.7B)
Method: LoRA SFT via Unsloth + TRL
Adapter config: r=16, lora_alpha=16, lora_dropout=0, targets q,k,v,o,gate,up,down
Training data: pkloats/bishop-chess-dataset (~18.6M tokens)
PEFT: 0.19.1

Files

File	What it is
`adapter_model.safetensors`, `adapter_config.json`	the LoRA adapter (apply on top of the base model)
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`	tokenizer + chat template
`qwen3-1.7b-bishop.Q4_K_M.gguf`	release GGUF (stage3c) — merged + quantized, runnable standalone
`checkpoints/*.Q4_K_M.gguf`	intermediate training-progression checkpoints (see below)

Use the GGUF (Ollama)

ollama run hf.co/pkloats/qwen3-1.7b-bishop

Or with llama.cpp:

llama-cli -hf pkloats/qwen3-1.7b-bishop --file qwen3-1.7b-bishop.Q4_K_M.gguf -p "..."

Training-progression checkpoints

The checkpoints/ folder holds the intermediate GGUFs from the training run, so you can compare stages:

File	Stage
`checkpoints/qwen3-1.7b-bishop.stage1.Q4_K_M.gguf`	stage 1
`checkpoints/qwen3-1.7b-bishop.stage2.Q4_K_M.gguf`	stage 2
`checkpoints/qwen3-1.7b-bishop.stage3a.Q4_K_M.gguf`	stage 3a
`checkpoints/qwen3-1.7b-bishop.stage3b.Q4_K_M.gguf`	stage 3b
`qwen3-1.7b-bishop.Q4_K_M.gguf` (root)	stage 3c — released

All checkpoints are the same Q4_K_M quant, so the bare ollama run hf.co/pkloats/qwen3-1.7b-bishop always resolves to the root release. To try a specific stage, download it directly and point a runtime at the file:

bash

hf download pkloats/qwen3-1.7b-bishop checkpoints/qwen3-1.7b-bishop.stage1.Q4_K_M.gguf --local-dir .
llama-cli --file checkpoints/qwen3-1.7b-bishop.stage1.Q4_K_M.gguf -p "..."

Use the LoRA adapter (PEFT)

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", device_map="auto")
model = PeftModel.from_pretrained(base, "pkloats/qwen3-1.7b-bishop")
tok = AutoTokenizer.from_pretrained("pkloats/qwen3-1.7b-bishop")

Intended use & limitations

Intended for chess study/analysis text generation, with emphasis on bishop-related ideas. It is a 1.7B model and will hallucinate illegal moves or incorrect evaluations; do not treat its output as an engine. Training data provenance is mixed (see the dataset card) — released for research use.

Framework versions

PEFT 0.19.1