librarian-curator-qwen35-9b-grooming

The Librarian — Curator "Grooming" LoRA for Qwen3.5-9B

A rank-8 LoRA adapter that teaches Qwen3.5-9B to run The Librarian's grooming job: periodically reviewing a slice of an agent's stored memories and improving it holistically — consolidating fragments into single documents, removing stale or duplicated entries, and deciding when the right action is to do nothing.

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 grooming 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 grooming cases compiled from the owner's own Librarian instance (private; not included) — deliberately small, by design. On the sealed evaluation cohort, this 100-case adapter beat 250- and 500-case siblings in blind review: the larger adapters had learned to merge more aggressively, which was precisely the failure being removed.
• Window: 8,192 tokens (prompt + completion), completion-only loss
• Note: trained on non-thinking completions — run with thinking disabled.

Results (sealed cohort, blind review)

• Exact gold matches: 33 → 38
• Prose drift: 43 cases → 35

Usage

PEFT / transformers

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-grooming")

llama.cpp (GGUF)

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

llama-server -m Qwen3.5-9B-Q8_0.gguf \
  --lora-scaled curator-qwen35-9b-grooming-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 grooming model config at this adapter (e.g. via the per-job model settings added in the curator unification, spec 043). Leave harvesting/intake on the base model or their own adapters. Grooming remains best run with its dry-run/proposal path — this adapter makes the decisions sharper, it does not make the job safe to run unattended.

License

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