QMSum locate-then-summarize: summarizer LoRA adapter (LFM2.5 1.2B)
A LoRA adapter for query-focused meeting summarization, trained on located transcript spans from
QMSum. It is stage 2 of a two-stage pipeline: a cross-encoder scores fixed-width windows of a
transcript against the query and packs the best ones to a word budget, then this adapter summarizes
only those spans. It was built to answer whether that decomposition lets a small model compete with
frontier models given the whole transcript, and on ROUGE it does.
This is part of the release for the paper Locate-then-summarize on QMSum: training regime and
architecture outweigh scale in query-focused multi-domain meeting summarization (Ertas AI, in
preparation).
The release
Model details
- Base model:
LiquidAI/LFM2.5-1.2B-Instruct, LFM Open License v1.0
- Adaptation: 4-bit QLoRA, LoRA r=16, alpha 32, dropout 0.05,
all-linear, lr 2e-4
- Trainable parameters: 11,108,352, which is 0.94% of the 1,181,448,960 total
- Download size:
adapter_model.safetensors is 44.46 MB; the full release directory with
tokenizer files is 49.30 MB
- Training data: 1,095 gold-span targets over QMSum's specific queries, 3 epochs, 207 optimizer
steps, effective batch 16, sequence length 6,144
- Hardware: one NVIDIA RTX 5070 Ti (16 GB), about five and a half GPU-hours for this stage plus
about 7 minutes for the locator
Results, full official QMSum test split
n=281, greedy decoding, one frozen scorer, a single test touch per configuration. The adapter is
the same in both rows; they differ in the locator and the retrieval budget in front of it. The
frozen protocol declares 900-word windows and a 3,000-word span budget; the promoted configuration
deviates on exactly those two locator-side values, and the paper reports both rows.
| Configuration | Locator | Budget | R1 | R2 | R-L | R-Lsum | BERTScore |
|---|
| Protocol-exact | MiniLM L6, 900-word windows (22.7M) | 3,000 words | 33.39 | 10.65 | 22.83 | 29.30 | 0.8680 |
| Deviating, promoted | MiniLM L12, 375-word windows (33M) | 2,000 words | 35.41 | 12.28 | 24.63 | 31.36 | 0.8733 |
The gap between these two rows is not demonstrably robust to reseeding: a reseeded run of the
deviating configuration scored below the protocol-exact one, and the training-seed range on this
benchmark is 1.59 ROUGE-1. Read the difference as a result about one training run.
Scoring: rouge-score==0.1.2, f-measure, Porter stemming on; bert-score==0.3.13 with
roberta-large. ROUGE-Lsum is computed over sentences segmented identically for candidate and
reference. These numbers are comparable to each other and to the other rows we generated. They
are not comparable to published QMSum numbers, which use different ROUGE implementations under a
protocol the benchmark never specified, because QMSum ships no evaluation code.
Latency: 2.629 s/query end to end at the protocol-exact settings above, of which the locator
is 0.022 s.
On-device, quantized export (measured 2026-08-04)
Q4_K_M base GGUF plus our f16 LoRA GGUF (exported adapter GGUF SHA256
5cf1ee8f659e39a63860d31b2546153fdad624fecd2718e38d532088bf9655e5), under llama.cpp b9351:
| Validation ROUGE-1 |
|---|
| bf16 pipeline | 35.39 |
| Q4_K_M export | 34.35 |
| difference | -1.04, 95% CI [-1.98, -0.11] |
Quote 34.35 for any on-device claim, never the bf16 figure. The quantization cost is real and
its interval excludes zero. Output length held steady (57.9 against 58.4 words) with zero empty
generations. Single-run latency was 0.39 s/query against 2.63 s for bf16, roughly 7x, indicative
only since it is one run.
How to load
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3
4base = AutoModelForCausalLM.from_pretrained(
5 "LiquidAI/LFM2.5-1.2B-Instruct", torch_dtype="bfloat16", device_map="auto")
6model = PeftModel.from_pretrained(base, "ErtasAI/qmsum-summarizer-lfm2.5-1.2b-lora")
7tokenizer = AutoTokenizer.from_pretrained("ErtasAI/qmsum-summarizer-lfm2.5-1.2b-lora")
The published numbers come from the release repo's evaluation harness, which builds the exact
prompt the adapter was trained on (query plus located spans) and decodes greedily. To reproduce
the table above, use that harness rather than a hand-rolled prompt.
Intended use
Query-focused summarization of meeting transcripts: given a transcript and a question about it,
produce a short answer grounded in the meeting. It expects to be given retrieved spans rather than
a whole transcript, which is the regime it was trained for.
Not evaluated for, and known limits
- General chat, factual QA, or any safety-critical use. It was trained on one task and scored on
one benchmark.
- Faithfulness. The reported metrics are token-overlap and embedding similarity. Neither
measures whether a summary states things the meeting did not. Concurrent work of ours evaluates
the same systems propositionally and is not part of this release.
- Whole transcripts. Neither stage reads an entire meeting. The pipeline retrieves 2,000 to
3,000 words against a median test transcript of 9,206 words.
- Domains outside QMSum, which covers academic (ICSI), product design (AMI) and parliamentary
committee meetings in English only.
- Reproducibility on 16 GB. Training sits at the edge of what the card holds: peak reserved
memory reaches 30.9 GB on a 17.1 GB card, so it runs in host-memory fallback. Of four training
seeds attempted, one completed, one exhausted memory at step 2, and two were abandoned.
Licence and what it obliges you to
The base model is under the LFM Open License v1.0 (copy in this repository as LICENSE), and
those terms travel with this adapter. This repository distributes a modification of
LFM2.5-1.2B-Instruct: a LoRA adapter produced by fine-tuning; the base weights themselves are
unmodified and are not redistributed here.
- Free commercial use is capped at US$10M annual revenue. Above that, you must contact Liquid
AI for a commercial licence. Qualified non-profits face no ceiling for research or
non-commercial use.
- Retain Liquid AI's copyright and licence notices, pass the licence on to anyone you distribute
to, and mark files you modify.
- There is no copyleft obligation.
This is a more restrictive licence than the rest of the artifacts in this release carry, and a
reader choosing between systems on our published scale is choosing between licences as well as
scores.
Attribution
Built on LFM2.5-1.2B-Instruct by Liquid AI, under the LFM Open License v1.0. Built on
QMSum (Zhong et al., NAACL 2021, MIT), which draws on the AMI and ICSI meeting corpora, both
released under CC BY 4.0 and both modified here into query-conditioned training pairs.
Provenance
| File | SHA256 |
|---|
adapter_model.safetensors | 75f871fa845ad90d02b5e2986269dd646faef1042000c796923b4d435e8759ed |
Released by
Ertas AI. We build custom small models that run on-device and
in your own infrastructure.
