A rank-8 LoRA adapter that teaches Qwen3.5-9B to run The Librarian's harvesting job: picking durable facts out of a settled conversation transcript and turning them into candidate memories worth keeping.

The Librarian is an open-source durable-memory layer for coding agents (code-ministry-ltd/the-librarian). Its curator runs three jobs — harvesting, intake, and grooming — each with its own job section of one unified prompt. This adapter was trained for the harvesting job only.

Base model

unsloth/Qwen3.5-9B (Apache-2.0)

Training

Method: SFT (TRL via Unsloth Studio), LoRA rank 8 / alpha 8, one epoch
Dataset: 100 harvesting cases compiled from the owner's own Librarian instance (private; not included)
Window: 8,192 tokens (prompt + completion), completion-only loss
Note: trained on non-thinking completions — run with thinking disabled.

Results (sealed cohort)

Harvesting F1 (joint precision/coverage): 78.2 → 80.8, improving both precision and coverage together rather than trading one for the other. Base and adapter returned the same 65 candidates, so the gain came from better selection, not from proposing more.

Usage

PEFT / transformers

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "unsloth/Qwen3.5-9B"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="cuda")
model = PeftModel.from_pretrained(model, "code-ministry-ltd/librarian-curator-qwen35-9b-harvesting")

llama.cpp (GGUF)

This repo also ships the adapter as a GGUF LoRA (curator-qwen35-9b-harvesting-100-r8-lora-f16.gguf, f16). Load it on any Qwen3.5-9B GGUF (Q8_0 recommended) — the scale is 1.0 by design:

bash

llama-server -m Qwen3.5-9B-Q8_0.gguf \
  --lora-scaled curator-qwen35-9b-harvesting-100-r8-lora-f16.gguf:1.0 \
  --jinja --chat-template-file <qwen3.5 chat template> \
  --chat-template-kwargs '{"enable_thinking": false}'

Notes for Librarian users

Point your curator's harvesting model config at this adapter (e.g. via the per-job model settings added in the curator unification, spec 043). Leave intake/grooming on the base model or their own adapters.

License

Adapter weights: Apache-2.0 (matching the base model). Training data was private and is not redistributed.