LiquidAI/LFM2.5-1.2B-Instruct fine-tuned with LoRA for QMD query
expansion using the v1-style verbose distillation 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-v1-verbose", torch_dtype=torch.bfloat16).
GGUF for QMD/llama.cpp:qmd-query-expansion-lfm2.5-1.2b-instruct-verbose-q5_k_m.gguf
The BF16 checkpoint is the unquantized merged fine-tune. The GGUF files are
derived from that same merged checkpoint.
This variant emits a deliberately verbose seven-line expansion:
one hyde:, three lex:, and three vec: lines. It was not trained for
Query intent: or /only:* 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.
The public historical query set was used as the source of
queries. Because it does not contain the unpublished labels used to train Tobil's
Qwen3 release, labels were reconstructed locally by teacher distillation from
tobil/qmd-query-expansion-1.7B
Q5_K_M. Accepted teacher outputs were quality-scored, deduplicated, and split
into 4,630 training and 516 untouched validation examples.
Teacher: tobil/qmd-query-expansion-1.7B Q5_K_M
Teacher recipe: 1 hyde + 3 lex + 3 vec; no intent or /only:* conditioning
Validation
Average QMD reward: 97.39%
Format compliance: 99.42%
Entity preservation: 100.00%
Hard failures: 0
Median QMD query-expansion latency: 0.959 s
p95 QMD query-expansion latency: 1.239 s
BF16 and GGUF comparison
Format
Avg reward
Δ vs BF16
Format compliance
Entity
BF16
97.70%
baseline
100.00%
100.00%
Q5_K_M
97.39%
-0.31 pp
99.42%
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 481.7 output tokens/s and 3.30 examples/s.
Q5_K_M through QMD: 0.959 s median and 1.239 s p95 query-expansion latency.
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.
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. The verbose v1-style variant also discloses its Tobil Qwen3 teacher provenance above.
