LFM2.5-230M-openjev

Merged BF16 weights of the LoRA decision fine-tune of LiquidAI/LFM2.5-230M-Base, trained so the model reliably answers a typed-decision prompt with a single label letter. It is the LFM member of the -openjev decision family, the small sibling of Qwen3-0.6B-openjev and Qwen2.5-1.5B-openjev.

Base model

This model was fine-tuned from:

LiquidAI/LFM2.5-230M-Base

Training

Supervised fine-tuning with LoRA, run with NeMo-RL, assistant-turn-masked cross-entropy over the typed-decision rows.

• LoRA: rank 32, alpha 32, dropout 0.05, target modules *.proj, *.w1, *.w2, *.w3 — the w1/w2/w3 names are LFM's feed-forward matrices, so the adapter reaches the MLP as well as the attention projections (a *.proj-only pattern would leave the MLP at base weights)
• Max sequence length: 4096
• Precision: bfloat16
• Optimizer: AdamW, lr 5e-5, weight decay 0.01
• Schedule: 100-step linear warmup, then cosine decay to 1e-6
• Global batch size: 64 (micro batch 2)
• Training: one epoch — 1191 steps over 76,224 rows (validation every 125)
• Loss masked to the assistant turn via the decision chat template
• SFT only, no GRPO: a decision is graded by the choice reward, and the SFT model is already at the ceiling on the rows GRPO samples, so the within-group advantage is zero and GRPO trains on no gradient here

The decision chat template is the model's contract: it renders the assistant turn as exactly its content (one label letter) with no thinking block, so the answer is the first generated token and the turn ends on EOS.

Dataset

Training used the train split of the ZefanCai/Open-Jev typed-decision corpus (snapshot c67699e13d0ae25e35b77165a4b6b079bedc8aba). Each row is a state, a question, a set of options, and a probability distribution over the options; the target is the argmax option, written as a single label letter. A row whose target ties (no argmax) is excluded where the loss is counted, not trained on a guess.

Intended use

• Typed-decision and classification tasks
• Single-label answer selection
• Small-footprint on-device or edge deployment

It is not intended to be a general replacement for larger models.

Sibling artifacts

• GGUF quantizations: iromu/LFM2.5-230M-openjev-GGUF
• NVFP4 (vLLM): iromu/LFM2.5-230M-openjev-NVFP4

Usage

Load a GGUF quant with llama.cpp:

llama-cli -hf iromu/LFM2.5-230M-openjev-GGUF:Q4_K_M

Load the merged BF16 model with transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "iromu/LFM2.5-230M-openjev", torch_dtype="bfloat16", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("iromu/LFM2.5-230M-openjev")

The chat template is the decision template: the answer is one letter and the turn ends on EOS, so a small max_new_tokens is more than enough — the model emits one token and stops.

Coherence

Probability-coherence score from JevBench, a metamorphic test of 50 laws of probability and choice with no gold labels (jevbench-mini suite: 240 cases, 50 scored relations, tolerance 0.05, temperature 1.0). Overall: 38.5 / 100.

• REP — Representation consistency: 60.6
• BAT — Batch independence: 25.8
• MEA — Probability measure coherence: 15.4
• LOG — Logical coherence: 68.3
• CHO — Choice-set coherence: 22.2

The score is dominated by MEA and CHO: the letter-emitter is overconfident and does not preserve the mean under transformation, so metamorphic coherence trails the Qwen3-0.6B-openjev sibling (44.5 on the same suite). Coherence and argmax accuracy measure different things — this model's argmax accuracy is high (77.3%) while its probability readout is miscalibrated.

Validation

The base model knows nothing of the decision format — it emits a lone label letter on 0% of rows (its reply is prose, never a single token the grader can read). The LoRA fine-tune teaches the format outright: it emits a lone letter on every row and is right on 77.3%.

Measured on the Open-Jev validation split (3,495 rows), greedy decoding. Throughput is single-stream greedy decode, not serving throughput.