GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
annieyii
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
annieyii
Model Tree
MediaTek-Research/Breeze-7B-Instruct-v1_0
Input Modalities
Output Modalities
Supported Functionality
Given a candidate explanation, one gold reason point, and that point's aspect, the adapter outputs 是 (asserted) or 否 (not asserted). Entailment is semantic: a paraphrase counts, a polarity reversal does not. Coverage for a case is the fraction of its reason points the candidate hits, which makes a score decomposable into the reasons it covered and the ones it missed.
The four aspects come from §5.2 of TIPO's Examination Guidelines on Likelihood of Confusion, restricted to factors decidable from the marks themselves:
| Aspect | Covers |
|---|---|
主要識別部分 | the salient element the two marks share |
差異 | the concrete distinguishing element |
外觀 | whether the marks resemble each other in appearance; examiners routinely write appearance, concept and pronunciation as one inseparable clause, and such a clause is kept whole under this aspect |
近似程度結論 | whether similarity is found, and to what degree |
Fed pure procedural boilerplate, this adapter reports coverage of 0.993, against 0.004 for a zero-shot Qwen2.5-7B-Instruct; LETITBE's qualification gate requires no higher than 0.05. Use it as a per-point classifier, not to produce a coverage score.
500 balanced held-out pairs, drawn from cases published after 2023 while training used cases up to 2023.
| value | |
|---|---|
| Accuracy | 0.988 |
| Precision | 0.977 |
| Recall | 1.000 |
| F1 | 0.988 |
| Specificity on polarity flips | ≥ 0.99 |
| Judgments that flip when the candidate is paraphrased | ≤ 1 in 100 |
Its precision is lower than the Gemma adapter's (0.977 against 0.992).
Use letitbe.py, which builds the exact prompt these numbers were measured with:
bash
python letitbe.py detector --input your_pairs.jsonl \ --backend hf --engine MediaTek-Research/Breeze-7B-Instruct-v1_0 --adapter annieyii/letitbe-hitdet-breeze-7bLoading it directly:
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigimport torch base_id = "MediaTek-Research/Breeze-7B-Instruct-v1_0"rev = "60d4fa07ae4ddd4174864b257be0563f1806cd6d" # the base revision the adapter was trained onbnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)tok = AutoTokenizer.from_pretrained(base_id, revision=rev, padding_side="left")base = AutoModelForCausalLM.from_pretrained(base_id, revision=rev, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(base, "annieyii/letitbe-hitdet-breeze-7b")It was trained and evaluated in 4-bit NF4 with double quantisation. Running it at another precision is untested.
LoRA r=16, α=32, dropout 0.05, on all seven linear projections (q,k,v,o,gate,up,down) of a 4-bit NF4 base. At most 10 epochs with early stopping on validation loss. 3,000 training and 500 validation examples, balanced 2:1:1 across positive, easy-negative and hard-negative, stratified by aspect and sampled deterministically.
Base revision: MediaTek-Research/Breeze-7B-Instruct-v1_0@60d4fa07ae4ddd4174864b257be0563f1806cd6d.
Versions: torch 2.10.0, transformers 5.12.1, peft 0.19.1, trl 1.7.0, datasets 5.0.0, tokenizers 0.22.2.
Apache-2.0, following its base MediaTek-Research/Breeze-7B-Instruct-v1_0. The full text is in LICENSE. The adapter files are new files produced by fine-tuning that base; no base-model file is redistributed here.
python letitbe.py detector --input your_pairs.jsonl \ --backend hf --engine MediaTek-Research/Breeze-7B-Instruct-v1_0 --adapter annieyii/letitbe-hitdet-breeze-7bfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigimport torch base_id = "MediaTek-Research/Breeze-7B-Instruct-v1_0"rev = "60d4fa07ae4ddd4174864b257be0563f1806cd6d" # the base revision the adapter was trained onbnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)tok = AutoTokenizer.from_pretrained(base_id, revision=rev, padding_side="left")base = AutoModelForCausalLM.from_pretrained(base_id, revision=rev, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(base, "annieyii/letitbe-hitdet-breeze-7b")