LiquidAI/LFM2.5-1.2B-Instruct fine-tuned with LoRA for QMD query expansion using the v2 data recipe. This repository contains the merged BF16 Transformers checkpoint at its root and QMD-ready GGUF quantizations alongside it.

Repository formats

Merged BF16 Transformers checkpoint: load the repository directly with AutoModelForCausalLM.from_pretrained("langleu/qmd-query-expansion-lfm2.5-1.2b-instruct-v2", torch_dtype=torch.bfloat16).
GGUF for QMD/llama.cpp: qmd-query-expansion-lfm2.5-1.2b-instruct-q5_k_m.gguf

The BF16 checkpoint is the unquantized merged fine-tune. The GGUF files are derived from that same merged checkpoint.

Artifact validated by this publishing invocation

File: qmd-query-expansion-lfm2.5-1.2b-instruct-q5_k_m.gguf
Kind: GGUF
Size: 843,354,528 bytes
SHA-256: 12cdc31cb1848cd1bd6b8dbf07c57339c5e2555550598e8469814837e5351c88
Base model: LiquidAI/LFM2.5-1.2B-Instruct
Base revision: 868df74dd56ff8a0c2ac5dbf281690c2dbebe4c9
Dataset: tobil/qmd-query-expansion-train-v2
Dataset revision: 9b9c3dd41a11209f3ce907908f9b865f0ead415e
Configured release quantizations: Q5_K_M
Evaluation report: q5_k_m

Prompt and behavior

/no_think Expand this search query: {query}

This variant supports the production v2 QMD behavior: an optional Query intent: line and /only:lex, /only:vec, and /only:hyde directives.

Each emitted line begins with lex:, vec:, or hyde:.

Use with QMD

The Transformers checkpoint is not loaded by QMD directly; select one of the GGUF files.

For a private repository, authenticate once with hf auth login, or set HF_TOKEN. QMD's node-llama-cpp downloader reads the cached Hugging Face token from ~/.cache/huggingface/token by default.

bash

export QMD_GENERATE_MODEL="hf:langleu/qmd-query-expansion-lfm2.5-1.2b-instruct-v2/qmd-query-expansion-lfm2.5-1.2b-instruct-q5_k_m.gguf"
qmd query --json --explain --no-rerank "docker container shutdown timeout"

Training-data provenance

Training records were rebuilt from the pinned dataset's messages field with the target LFM2.5 tokenizer. The dataset's preformatted Qwen text was not used. Completion-only loss masked the user prompt.

Validation

Average QMD reward: 94.42%
Format compliance: 99.68%
Entity preservation: 96.61%
/only:* behavior: 100.00%
Hard failures: 0

BF16 and GGUF comparison

Format	Avg reward	Δ vs BF16	Format compliance	Entity	`/only:*`
BF16	94.47%	baseline	100.00%	96.61%	100.00%
Q5_K_M	94.42%	-0.05 pp	99.68%	96.61%	100.00%

Quality changes are reported in percentage points (pp), not relative percent. Small positive GGUF deltas can occur because sampled generation is not bit-for-bit deterministic.

BF16 (batched Transformers): approximately 406.9 output tokens/s and 4.04 examples/s.
Q5_K_M (GGUF evaluation): approximately 126.1 output tokens/s and 1.28 examples/s.

Speed percentages are intentionally not reported. BF16 was measured with batched Transformers inference, while QMD runs GGUF through llama.cpp one query at a time. Their observed throughput and latency are useful operational measurements, but dividing them would not be an apples-to-apples speedup.

Credits

Liquid AI for LFM2.5 and the LFM Open License v1.0.
Tobi for QMD, the QMD query-expansion datasets, evaluation/scoring design, and the Qwen3 query-expansion model.
OrcsRise for the earlier LFM2 QMD fine-tuning work that informed the LFM target-module recipe.
QMD, TRL, PEFT, and llama.cpp.

Licensing and dataset notice

This derivative is governed by the LFM Open License v1.0. The included LICENSE must be retained, including its attribution and commercial-use terms.

The upstream QMD dataset card did not declare an explicit dataset license at the pinned revision. This repository records that fact and does not imply that a license was granted. Users and redistributors are responsible for confirming that their use is authorized.