Structured output remains a stubborn bottleneck for LLMs. Most benchmarks bury this capability inside broader reasoning or extraction scores, yet schema compliance is often the deciding factor for whether a model can actually plug into a downstream system.
This guide covers a specific training pipeline. It does not attempt to recreate the IFStruct benchmark score from the IFStruct blog. The goal is to show how task-specific fine-tuning of smaller models can improve performance and match that of far larger models.
Prerequisites
The workflow splits into two environments:
- Fine-tuning runs on a GPU. The accompanying notebook fits on a free-tier Colab or Kaggle GPU.
- Evaluation runs locally on a MacBook. This example uses a MacBook Pro with an Apple M5 Max and 36 GB of unified memory through
llama.cpp
. This tool exposes an OpenAI-compatible server that the IFStruct evaluator talks to.
You need
uv
for Python tooling and
llama.cpp
for serving. Follow the Liquid AI llama.cpp deployment docs to install
llama.cpp
with Homebrew and verify
llama-server
is available:
brew install llama.cpp llama-server --versionIFStruct Evaluation on LFM2.5-350M (Base model)
We begin by evaluating LFM2.5-350M on the IFStruct benchmark to see if we can reproduce the reported score of 21.1%.
IFStruct tests the validity of LLM outputs and schema adherence. The benchmark is open-source in Liquid4All/ifstruct, with the public dataset on Hugging Face at LiquidAI/ifstruct-v1.0.
git clone https://github.com/Liquid4All/ifstruct.gitFor the comparison, we serve the model locally on the MacBook using
llama.cpp. We use the
BF16GGUF file from LiquidAI/LFM2.5-350M-GGUF.
Start the base-model server with this command:
llama-server \ -hf LiquidAI/LFM2.5-350M-GGUF:BF16 \ -c 32768 \ -np 4 \ -ngl 99 \ --alias LiquidAI/LFM2.5-350M \ --host 127.0.0.1 \ --port 8080
--alias
: the model name IFStruct sends to the OpenAI-compatible endpoint
-ngl 99
: asks
llama.cpp
to offload all layers to the GPU when available
-np 4
: serves four requests in parallel
-c 32768
: size of the prompt context
Once the server is running, run the full benchmark with 2000 samples:
uv run ifstruct-eval \ --model LiquidAI/LFM2.5-350M \ --base-url http://localhost:8080/v1 \ --api-key dummy \ --dataset data/test.jsonl \ --results-file results/lfm2.5-350m-llamacpp-base.json \ --n-threads 4 \ --max-tokens 2048 \ -v============================================================ Model: LiquidAI/LFM2.5-350M ============================================================ Overall: 452/2000 passed (22.6%) Average latency: 1453ms By format: JSON: 180/1000 passed (18.0%) YAML: 272/1000 passed (27.2%) By top-level structure: Wrapper key 288/1011 passed (28.5%) Bare list 164/989 passed (16.6%) By entity type: test__camera_review 6/83 passed (7.2%) test__clinical_trial 20/104 passed (19.2%) test__conference_schedule 7/87 passed (8.0%) test__escaping__bug_report_batch 24/89 passed (27.0%) test__escaping__config_snippet_audit 15/85 passed (17.6%) test__escaping__customer_email_thread 5/73 passed (6.8%) test__escaping__dialogue_sample 14/95 passed (14.7%) test__escaping__interview_transcript_segment 21/80 passed (26.2%) test__escaping__log_parser_examples 21/72 passed (29.2%) test__escaping__pr_discussion 22/87 passed (25.3%) test__escaping__repro_steps_batch 16/73 passed (21.9%) test__escaping__screenplay_scene 16/92 passed (17.4%) test__escaping__short_story_chapter 15/84 passed (17.9%) test__escaping__support_ticket_batch 27/73 passed (37.0%) test__escaping__terminal_session_notes 20/70 passed (28.6%) test__event_ticket_booking 49/107 passed (45.8%) test__gpu_review 6/94 passed (6.4%) test__invoice 28/86 passed (32.6%) test__job_posting 25/85 passed (29.4%) test__real_estate_listing 31/82 passed (37.8%) test__recipe 3/70 passed (4.3%) test__rental_car_booking 27/79 passed (34.2%) test__scientific_experiment 13/69 passed (18.8%) test__travel_itinerary 21/81 passed (25.9%) Common errors: 7228x required field missing 738x wrong item count 540x type mismatch 317x Unclosed code block 190x extraneous field 'notes' 181x extraneous field 'path' 175x extraneous field 'constraints' 170x extraneous field 'type' 170x missing code block 100x expected bare list, got wrapperThe IFStruct release blog reports 21.1% for LFM2.5-350M. Our local llama.cpp/BF16 setup measures 22.6%, close to the 21.1% reported in the IFStruct blog. We use this local result as the baseline for the same serving stack comparison.
GRPO Fine-tuning with TRL on Structured Outputs
The full, runnable pipeline lives in the accompanying notebook. We cover only the relevant pieces here.
Training data
We use
nvidia/Nemotron-RL-instruction_following-structured_outputs. This pairs each prompt with a target JSON Schema and an expected field count. We use about 500 samples for training.
Because the Nemotron data distribution differs from the IFStruct evaluation, we augment the prompts to close two gaps:
- 40% get a “return the output inside a fenced code block” instruction appended. This teaches the model to follow the format instruction rather than always emitting raw JSON.
- A disjoint 20% are converted into top-level-array tasks. The schema is wrapped in an
array
with a required item count, which trains bare-list output and item-count compliance.
Model and LoRA
We load
LiquidAI/LFM2.5-350M
and attach a LoRA adapter. Because LFM2.5 uses a hybrid attention/convolution architecture, we target the LFM-specific module names:
lora_config = LoraConfig( r=16, lora_alpha=32, bias="none", task_type="CAUSAL_LM", target_modules=[ "q_proj", "k_proj", "v_proj", "out_proj", "in_proj", "w1", "w2", "w3", ], )This trains approximately 6M parameters, about 1.66% of the model.
Reward functions
Three reward functions score every completion on whether the extracted structure is correct. Each runs on a
[0, 1]scale:
json_format_reward
: Is the output parseable and in the requested form? Full credit (
1.0
) for the requested form (fenced vs. raw),
0.2
for the wrong-but-parseable form,
0.0
for unparseable output.
field_count_reward
: Does the object have the expected number of top-level fields? An exact match earns
1.0
, and the score decays linearly with the miss.
schema_validation_reward
: Does the output validate against the row’s JSON Schema? It counts every constraint violation and gates partial credit on required-key coverage.
We combine the three as a weighted sum with
reward_weights=[1.0, 0.5, 2.0]
.
Training
We train for 100 steps with 8 generations per prompt group, sized for a free-tier 16 GB GPU:
from trl import GRPOConfigtraining_args = GRPOConfig(
output_dir="./outputs/lfm25-350m-nemotron-schema-grpo",
learning_rate=5e-5,
max_steps=100,Source Read original →
