Downloads · 30 days
18
15% of all-time downloads
banhchungtuongot/qwen25-1p5b-massive-orpo-boundary
qwen25-1p5b-massive-orpo-boundary is a machine learning model from banhchungtuongot. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
Slot-boundary preference-aligned LoRA adapters for Amazon MASSIVE (en-US + vi-VN) intent + slot extraction, on top of Qwen/Qwen2.5-1.5B-Instruct. This single repo bundles the trained adapters and the self-mined prefer…
Downloads · 30 days
18
15% of all-time downloads
All-time downloads
119
Public
Repo size
602 MB
Likes
0
Public
Click a slice to open those files.
.safetensors591 MB · 95%
From the Hugging Face model README
Slot-boundary preference-aligned LoRA adapters for Amazon MASSIVE (en-US + vi-VN)
intent + slot extraction, on top of Qwen/Qwen2.5-1.5B-Instruct. This single repo
bundles the trained adapters and the self-mined preference data used to make them.
| Path | What |
|---|---|
./ (root) | ORPO keeper adapter — load directly with PeftModel.from_pretrained(base, "<repo>") |
dpo_boundary_adapter/ | DPO variant (Experiment 1) |
preference_data/massive_pref_orpo_*.jsonl | 1064 self-mined ORPO preference pairs (769 vi + 295 en) |
preference_data/massive_boundary_dpo_*.jsonl | DPO preference pairs (Experiment 1) |
Both adapters continue the SFT-LoRA MASSIVE adapter
banhchungtuongot/small-conv-llm-massive-h100-e
with hard negatives where chosen = gold target and rejected = the model's own
boundary / over-segmentation / wrong-key error. Pairs were mined from the TRAIN+DEV
splits only (no test leakage).
| Metric | h100_e baseline | DPO | ORPO (root, keeper) |
|---|---|---|---|
| overall slot-F1 | 0.7770 | 0.7795 | 0.7824 (+0.54) |
| overall exact match | 0.6982 | 0.7024 | 0.7017 (+0.35) |
| vi-VN slot-F1 | 0.7439 | 0.7443 | 0.7471 (+0.32) |
| en-US slot-F1 | 0.8101 | 0.8146 | 0.8177 (+0.76) |
| json parse rate | 1.0000 | 1.0000 | 1.0000 (no collapse) |
Net-positive, no regression, no JSON collapse. Method: ORPO (trl 0.19.1), β=0.2,
lr 8e-6, 1 epoch. Gains are modest by design — the residual vi-VN gap is largely a
label-noise ceiling (MASSIVE-VI classifier/MT-artifact inconsistency); scaling the
pair set further regressed the model. Full analysis: docs/error-patterns.md in the
source repo.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "banhchungtuongot/qwen25-1p5b-massive-orpo-boundary")
# DPO variant: PeftModel.from_pretrained(model, "<repo>", subfolder="dpo_boundary_adapter")
msgs = [
{"role": "system", "content": "Return valid JSON only with keys intent, slots, missing_slots, next_action."},
{"role": "user", "content": "Locale: vi-VN\nUtterance: đánh thức tôi dậy lúc năm giờ sáng tuần này"},
]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=128, do_sample=False)[0][ids.shape[1]:], skip_special_tokens=True))