Modelpublic

kader-xai/priya-qwen2.5-7b-lora

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes8downloads

Model Card

Priya — Qwen 2.5 7B LoRA adapter

A LoRA adapter fine-tuned on top of `Qwen/Qwen2.5-7B-Instruct` to draft email replies in the voice of Priya Sharma, a (fictional) Senior Customer Success Manager at a synthetic SaaS company. Built as part of Project Recall — an experiment in capturing a departing employee's "institutional memory" as a draft-only AI assistant.

The persona is fictional. All training data is synthetic. Nothing in this adapter encodes real person, customer, or company information.

—GGUF for local use: `kader-xai/priya-qwen2.5-7b-gguf`
—Training dataset: `kader-xai/priya-sft`
—📝 Blog post: Employee Recall — Capturing a Departing Employee's Writing Style and Memory
—Source code + methodology: github.com/kader-xai/EmployeeRecall
—How to train your own persona: train-own-persona.md

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct", torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(base, "kader-xai/priya-qwen2.5-7b-lora")
tokenizer = AutoTokenizer.from_pretrained("kader-xai/priya-qwen2.5-7b-lora")

system_prompt = open("system_prompt_priya.txt").read()  # from the dataset repo
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Customer asking when the v2 dashboard ships. Draft a reply."},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=400, temperature=0.4, top_p=0.9)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

For grounded answers ("what happened with Account X?"), pair the model with the RAG index built from kader-xai/priya-sft. See the training article for the retrieval + prompt-composition pattern.

Training details


Base model	`Qwen/Qwen2.5-7B-Instruct` (loaded as `unsloth/Qwen2.5-7B-Instruct-bnb-4bit`)
Method	LoRA via PEFT + TRL SFTTrainer, accelerated with Unsloth
LoRA rank / alpha	16 / 32
LoRA dropout	0.05
Target modules	`q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj`
Max seq length	4096
Epochs	3
Effective batch	16 (4 × grad-accum 4)
Learning rate	2e-4, cosine, warmup 0.03
Precision	4-bit base + bf16 adapter
Hardware	1× A100 40GB
Dataset	1,223 train / 64 eval SFT pairs (`kader-xai/priya-sft`)

Eval

Metric	Value
Style cosine (BGE) vs held-out replies	0.735
History keyword recall (RAG-grounded)	0.30 on bulk corpus, 1.0 on hand-written storyline questions

The history score is bounded by corpus shape: questions whose answers exist only in templated bulk content can't be answered when that content doesn't carry the underlying fact.

Intended use & limitations

—Intended: demos, research on persona modeling, draft-only assistants with a human in the loop.
—Not intended: auto-sending replies, impersonating real people, factual Q&A without RAG grounding.
—Bias: the persona's voice fingerprint is hand-crafted and culturally specific (US-based SaaS CSM). Outputs reflect that.

License

Apache-2.0, inherited from the base model.