HoangHa/lfm2.5-230m-title-v0

Meddies Title SFT v0

Meddies Title SFT v0 is a private research candidate for generating short sidebar titles from user queries. It is a PEFT LoRA adapter for LiquidAI/LFM2.5-230M, trained on title examples in 17 languages.

Status: private, internal research. This is not a production release.

Intended use

Use this adapter only to generate one short session title from a query, with the same system and user prompt contract used during training. The 17 represented languages are:

German, English, Spanish, Filipino, French, Indonesian, Japanese, Korean, Lao, Malay, Burmese, Portuguese, Russian, Tamil, Thai, Vietnamese, and Chinese.

It is not a general assistant, clinical decision model, factual answerer, safety classifier, or replacement for application-level title validation.

Loading the adapter

Access to both this private repository and the base model is required.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "LiquidAI/LFM2.5-230M"
adapter_id = "Meddies/lfm2.5-230m-title-v0"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

def generate_title(messages):
    """Generate from messages built with the frozen Meddies title prompt."""
    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True,
    )
    inputs = {key: value.to(model.device) for key, value in inputs.items()}
    with torch.inference_mode():
        output = model.generate(**inputs, max_new_tokens=32, do_sample=False)
    return tokenizer.decode(
        output[0, inputs["input_ids"].shape[1]:],
        skip_special_tokens=True,
    ).strip()

Build messages with the versioned system prompt and language-specific user prompt from the Meddies title pipeline; do not substitute a generic instruction when comparing this candidate with recorded results. The example preserves and passes the tokenizer-produced attention_mask during generation.

Training

Item	Recorded value
Base model revision	40cb2ad3b3044d5a41eee083a6103c8b523afa45
Objective	Response-only causal-language-model SFT
Training rows	68,961
Validation rows	3,639
Languages	17
Epochs	1 completed and certified
Optimizer steps	4,311
Useful non-masked target tokens	1,687,393
Maximum sequence length	8,192 tokens
Learning rate	0.0004
Per-device batch size	2
Gradient accumulation	4
Seed	3407
LoRA rank / alpha	128 / 256
LoRA targets	All linear modules selected by the training stack
Checkpoint interval / retention	500 optimizer steps / at most 3 checkpoints

The training run used transformers==5.5.0, peft==0.19.1, torch==2.10.0, and unsloth==2026.5.2. The finalization receipt reports two Tesla T4 GPUs. The source SFT dataset was HoangHa/meddies-title, configuration title_sft_v0, pinned to revision 13ec07308f8bdcd0ababf66513d829beea9fa097.

The split and post-truncation guards checked that training, validation, and reserved benchmark queries did not collide. Twenty-four training queries and one validation query required middle truncation under the 8,192-token limit.

Evaluation

The completed checkpoint was evaluated on the full 3,639-row validation split:

• Validation loss: 1.254838228225708
• Evaluation throughput: 10.271 samples/second

A separate 3,400-group benchmark was reserved by manifest, but no benchmark outcome is reported here. Validation loss measures the held-out SFT objective; it is not evidence of title usefulness, sidebar fit, factual fidelity, language quality, or production readiness.

Checkpoint-time qualitative diagnostics emitted an attention-mask warning because the padding token equaled the end-of-sequence token. Therefore, this card makes no qualitative claim from those diagnostic generations. Any qualitative evaluation must be rerun with an explicit attention mask and the frozen prompt contract.

Limitations

• The adapter is specialized for title generation and may behave unpredictably outside that task or outside its training prompt format.
• Coverage across 17 languages does not establish equal quality across languages, scripts, domains, or query lengths.
• The recorded validation loss is not a user-facing quality metric.
• The reserved benchmark has not been reported in this card.
• Application code must still enforce language, grapheme-length, privacy, and output-format constraints.
• This candidate has not been established as safe for clinical decisions or as a source of medical facts.

Reproducibility and provenance

Artifact	SHA-256
adapter_model.safetensors	c814264e190f10322df63f4ba859ca4d1b53a4923c435c6e6d9591f969563213
adapter_config.json	7adab715ac3bd3bd38117e5f8301d6574ef618b8a3a753ee6782162ac204d3ea
SFT payload	480df3973bb2e7d719830864d8baac8f60e0a09cab6e59b92bacfd89ad55a914
Group-safe split manifest	fd3a91fb61414ea09df48f43e3227d3276ba0bf11f991a8792630c77b43c5302
Benchmark manifest	8fec57443c1af3e07dd472d4ebdde57148c69e4f1b0180d83c995095d07f25c6
Benchmark case IDs	73fd5d8ca4640c333fbad7fa8f6bb9c62b4edd4930c2da2fdcf7f1a8feec9de5
Training-trajectory train.py	034d6fb3cf80c90edb8294f96c2dbd56da10f67aadeff15e8b21b70e3d30dfb8
Finalization train.py	0b99ca40b9e76477ab18d90d7b41d241f511ed4f8ed0bab265f7957874c881d7

The finalization receipt records global step 4311 of expected step 4311, effective epoch 1.0, completed_one_epoch: true, and certified: true.