Lore

Lore is a coding-focused QLoRA fine-tune of Qwen/Qwen2.5-Coder-7B-Instruct that reached 56.4% pass@2 on 133 old-aider Python tasks. It was trained end to end on a single 8 GB consumer GPU and is distributed as an 8.1 GB Q8_0 GGUF for local inference.

Project Highlights

• Built an end-to-end 4-bit QLoRA pipeline with Unsloth, TRL, and PEFT, including rank-16 LoRA, gradient checkpointing, checkpoint recovery, and adapter continuation on a single 8 GB RTX 3060 Ti.
• Created execution-verified Python repair data with deterministic AST mutation, isolated subprocess testing, tokenizer-aware limits, and parent-disjoint train/validation splits.
• Reached a 56.4% pass@2 high score on 133 old-aider Python tasks, above the published leaderboard entries for Qwen2.5-Coder 7B Q8_0 (51.9%), Claude 3 Sonnet (54.9%), and GPT-4o mini (55.6%).
• Exported the final model as an 8.1 GB Q8_0 GGUF for local Ollama and llama.cpp inference.

Model Details

• Format: GGUF
• Quantization: Q8_0
• File size: 8,098,525,408 bytes
• SHA-256: d9a86d7f85b433f3f4bca84aa150c6db001133907d93446d7b116b886c512655
• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Initial fine-tuning: bounded 2,000-example subset of Nexlab/fable5-agentic-coding-sft
• Final continuation: 128 execution-verified repair examples formatted as multi-turn retry conversations
• Final continuation context length: 1,024 tokens
• Final continuation steps: 16
• Final continuation learning rate: 5e-7

Evaluation

Lore was evaluated twice with aider v0.56.0 on the 133-task Exercism Python benchmark, using whole-file edit format and up to two attempts.

Run	Pass rate 1	Pass rate 2
First	48.1%	56.4% (75/133)
Confirmation	46.6%	54.9% (73/133)

The two-run pass@2 mean was 55.64%. The public leaderboard comparisons above use the same old-aider benchmark family, but local runtime and harness details may differ. Results also showed task-level variance, so the single-run high score should not be treated as a stable estimate across other prompts, inference settings, or benchmarks.

Usage

Ollama

Download this repository, then create the model with the included Modelfile:

ollama create lore -f Modelfile
ollama run lore

llama.cpp

llama-cli \
  -m qwen2.5-coder-7b-instruct.Q8_0.gguf \
  -cnv \
  -p "Write a Python function that checks whether a string is a palindrome."

Limitations

This model is specialized for Python coding-agent workflows and complete-file responses. It may produce incorrect, insecure, or incomplete code. Review and test generated code before use. The GGUF is quantized and may not exactly match the unquantized adapter's behavior.