Model Card for codenopro/lab22-dpo-vn

A LoRA adapter that DPO-aligns Qwen2.5-3B for Vietnamese instruction following. Pipeline: SFT-mini → DPO (the DPO LoRA continues training the SFT LoRA, so this single adapter is the final SFT+DPO model). Built for Day 22 / Track 3 (DPO/ORPO Alignment) of the VinUni AICB program on a free Colab T4.

Model Details

Model Description

This adapter was produced by (1) a small supervised fine-tune of unsloth/Qwen2.5-3B-bnb-4bit on a 1,000-sample Vietnamese Alpaca slice, then (2) Direct Preference Optimization (TRL DPOTrainer, β=0.1) on 2,000 binarized UltraFeedback preference pairs. It is a teaching-scale run: DPO shifted the implicit-reward margin but produced only marginal, noisy behavior changes (see Evaluation). It is not a production or safety-aligned model.

• Developed by: codenopro (VinUni AICB student)
• Funded by [optional]: N/A — free Google Colab T4
• Shared by [optional]: codenopro
• Model type: PEFT LoRA adapter for a decoder-only causal LM (DPO-aligned)
• Language(s) (NLP): Vietnamese (vi)
• License: apache-2.0 (inherits Qwen2.5-3B)
• Finetuned from model [optional]: unsloth/Qwen2.5-3B-bnb-4bit

Model Sources [optional]

• Repository: https://huggingface.co/codenopro/lab22-dpo-vn — lab code: https://github.com/anhkiet75/Day22-Track3-DPO-Alignment-Lab

Uses

Direct Use

Vietnamese instruction following / chat using the Qwen2.5 ChatML template. Intended for research and education on the SFT→DPO alignment pipeline.

Downstream Use [optional]

A starting point for further preference-tuning experiments (e.g. β-sweeps, ORPO/SimPO comparisons) or as a worked example of LoRA DPO on a 4-bit base.

Out-of-Scope Use

Not for production, user-facing assistants, factual/medical/legal advice, or any safety-sensitive setting. The model is not safety-aligned (see Limitations) and degenerates on a small SFT base.

Bias, Risks, and Limitations

• Small-base ceiling: the 1,000-sample SFT base degenerates — repetition loops, occasional drift into English, and stray "turn-leak" tokens. Both SFT-only and SFT+DPO inherit this.
• Safety not solved: on adversarial prompts (home explosives, terror message) both the base and this adapter partially or fully comply. DPO improved exactly one safety refusal and left the rest flat.
• Marginal alignment effect: at 3B / 4-bit / 250 DPO steps, effects are small and noisy. The reward gap rose during training while generations barely changed — a textbook "rising margin ≠ better outputs" case.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Do not deploy it where unsafe or low-quality completions could cause harm; treat outputs as illustrative, not reliable.

How to Get Started with the Model

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

base = "Qwen/Qwen2.5-3B"  # or load unsloth/Qwen2.5-3B-bnb-4bit in 4-bit
tok = AutoTokenizer.from_pretrained("codenopro/lab22-dpo-vn")
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "codenopro/lab22-dpo-vn")

msgs = [{"role": "user", "content": "Giải thích ngắn gọn cách thuật toán quicksort hoạt động."}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Training Details

Training Data

• SFT: 5CD-AI/Vietnamese-alpaca-gpt4-gg-translated — 1,000-sample slice, 1 epoch.
• Preference (DPO): argilla/ultrafeedback-binarized-preferences-cleaned — 2,000 (prompt, chosen, rejected) pairs, 1 epoch.

Training Procedure

Two LoRA phases on the same adapter: SFT (NB1), then DPO (NB3) with the reference model auto-derived from the PEFT base (adapter disabled for the reference forward pass — no second copy of weights).

Preprocessing [optional]

Examples formatted to Qwen2.5 ChatML (<|im_start|> / <|im_end|>) via tokenizer.apply_chat_template. Max sequence length 512, max prompt length 256.

Training Hyperparameters

• Training regime: fp16 (Turing T4), 4-bit base (bitsandbytes NF4) + LoRA
• LoRA: r=16, lora_alpha=32, dropout 0, target q,k,v,o,gate,up,down_proj
• SFT: 1 epoch (125 steps), batch 1 × grad-accum 8, loss ≈ 1.87 → ~1.5
• DPO: beta=0.1, loss_type=sigmoid, lr 5e-7, 250 steps (1 epoch), effective batch 8

Speeds, Sizes, Times [optional]

DPO took ≈ 40 min on a free T4 — slower than usual because T4 (compute 7.5) cannot run xformers' grouped-query-attention backward, so a PyTorch SDPA math-backend attention fallback was used. Final DPO loss 0.7719; end reward gap (chosen − rejected) ≈ 0.14 (last step) / ≈ 0.20 (last-5 mean). Adapter weights only (LoRA), not full model.

Evaluation

Testing Data, Factors & Metrics

Testing Data

8 held-out Vietnamese prompts (4 helpfulness + 4 safety) — data/eval/side_by_side.jsonl in the lab repo.

Factors

Disaggregated by prompt category: helpfulness vs safety.

Metrics

Manual win/loss/tie between SFT-only and SFT+DPO on full generations (no API judge).

Results

Metric	Result
SFT+DPO vs SFT-only	1 win / 1 loss / 6 ties
Helpfulness	0W / 1L / 3T
Safety	1W / 0L / 3T
Final DPO loss	0.7719
End reward gap	≈ 0.14

Summary

DPO's one clear win is a cleaner safety refusal (underage-alcohol prompt); its one loss is a repetition collapse on a quicksort explanation where SFT-only stayed coherent. Net effect is essentially a wash — consistent with the small base and short training.

Model Examination [optional]

The reward gap rose steadily while the qualitative win-rate stayed flat (1–1–6). This is the classic DPO caveat: optimizing the implicit-reward margin (chosen − rejected) does not guarantee better generations, and can even introduce new failure modes (more repetition) on a weak base.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA Tesla T4 (16 GB)
• Hours used: ≈ 1 hour total (SFT ~7 min + DPO ~40 min + eval)
• Cloud Provider: Google Colab (free tier)
• Compute Region: Unknown (Colab-assigned)
• Carbon Emitted: Not measured; negligible (single T4, < 1 GPU-hour)

Technical Specifications [optional]

Model Architecture and Objective

Qwen2.5-3B decoder-only causal LM with a LoRA adapter; objective is the DPO sigmoid loss over preference pairs (with an SFT cross-entropy warm-up phase).

Compute Infrastructure

Hardware

Single NVIDIA Tesla T4 (16 GB), 4-bit quantized base + LoRA.

Software

Unsloth 2026.4.8 · TRL · PEFT 0.19.1 · Transformers 5.5.0 · PyTorch 2.10.0+cu128 (CUDA 12.8) · bitsandbytes.

Citation [optional]

Day 22 · Track 3 · VinUni AICB — DPO/ORPO Alignment lab. Base model: Qwen2.5-3B (Apache-2.0). Trained with Unsloth + TRL.

BibTeX:

N/A (course lab; no associated paper)

APA:

codenopro. (2026). lab22-dpo-vn: SFT→DPO LoRA adapter for Qwen2.5-3B (Vietnamese). Hugging Face. https://huggingface.co/codenopro/lab22-dpo-vn

Glossary [optional]

• DPO — Direct Preference Optimization: trains directly on (chosen, rejected) pairs using the policy/reference log-prob ratio, no separate reward model.
• Reward gap — chosen − rejected implicit reward; the headline DPO diagnostic.
• LoRA — low-rank adapters injected into attention/MLP projections; only these are trained.

More Information [optional]

See the lab repository for the full pipeline, reward-curve plot, side-by-side comparison, and reflection: https://github.com/anhkiet75/Day22-Track3-DPO-Alignment-Lab

Model Card Authors [optional]

codenopro

Model Card Contact

Via the Hugging Face repository: https://huggingface.co/codenopro/lab22-dpo-vn

Framework versions

• PEFT 0.19.1
• TRL · Transformers 5.5.0 · PyTorch 2.10.0+cu128 · Unsloth 2026.4.8