PresenceAI

Polyhymnia-32B-SFT-LoRA-v1.0

README

License: apache-2.0

Model details

Table with columns: Item, Value
Item	Value
Developed by	PresenceAI
Series	Presence Muses
Release	v1.0 (GA)
Model type	LoRA adapter (PEFT)
Training stage	Supervised fine-tuning (SFT)
Base model	`Qwen/Qwen3-32B`
Languages	English, Chinese
License	Apache-2.0
Adapter rank	r = 32, alpha = 32, dropout = 0.0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`

Model lineage

text

Qwen/Qwen3-32B  ──SFT (full data)──▶  Polyhymnia-32B-SFT-LoRA-v1.0  (this repository)

Trained directly from the base model. No intermediate checkpoint or upstream adapter is required — load this adapter on Qwen/Qwen3-32B as shown below. The base weights are not included in this repository.

Intended use

Multi-turn roleplay and companion-style conversation.
Tool-calling assistants that emit a send_message call.
Research on persona consistency and conversational style adaptation.

Out of scope: factual question answering, medical / legal / financial advice, high-stakes decision making.

Training

Method: supervised fine-tuning (SFT) with LoRA.
Training data: 129,482 internal multi-turn conversations with tool-call annotations. The training data is not released with this adapter.
Maximum sequence length: 22,528 tokens.
Epochs: 1.
Effective batch size: 16 (per-device batch 1 × gradient accumulation 16).
Learning rate: 1e-4.
LoRA: r = 32, alpha = 32, dropout = 0.

Evaluation

Automatic evaluation on an internal held-out split (token-level eval loss):

Table with columns: Step, Eval loss, Note
Step	Eval loss	Note
4,200	0.6023	Best value recorded in the training log
4,300	0.6029
4,500	0.6058

The training run continued beyond the recorded evaluation history. The value above is the best recorded during training and was not re-measured on the exact released checkpoint. No task-level benchmark, human preference study, or safety evaluation is included.

Usage

PEFT / Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-32B"adapter_id = "PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0" tokenizer = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(    base_id, torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, adapter_id)model.eval()

vLLM

bash

vllm serve Qwen/Qwen3-32B \  --enable-lora --max-lora-rank 32 \  --lora-modules polyhymnia=PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0 \  --enable-auto-tool-choice --tool-call-parser hermes

Tool-call format used during training:

text

<tool_call>{"name": "send_message", "arguments": {"text": "..."}}</tool_call>

Use the tokenizer and chat_template.jinja included in this repository.

Limitations

Trained for roleplay/companion conversation; may be inappropriate outside that domain.
May reproduce patterns or content from its training data. The training data is not released; deploy with privacy and safety filters.
Only automatic eval loss is reported. No factual-accuracy, safety, or human-quality claim is made.
The base model's limitations and license terms continue to apply.

License

The adapter is released under the Apache License 2.0. The base model Qwen/Qwen3-32B is also Apache-2.0. See LICENSE and NOTICE.

Citation

bibtex

@misc{presenceai_polyhymnia_32b_sft_lora_v1,  title  = {PresenceAI Polyhymnia-32B-SFT-LoRA v1.0: roleplay and tool-calling chat},  author = {PresenceAI},  year   = {2026},  note   = {LoRA adapter, Apache-2.0}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

PresenceAI

Model Tree

Base

Qwen/Qwen3-32B

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model details

Table with columns: Item, Value
Item	Value
Developed by	PresenceAI
Series	Presence Muses
Release	v1.0 (GA)
Model type	LoRA adapter (PEFT)
Training stage	Supervised fine-tuning (SFT)
Base model	`Qwen/Qwen3-32B`
Languages	English, Chinese
License	Apache-2.0
Adapter rank	r = 32, alpha = 32, dropout = 0.0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`

Model lineage

text

Qwen/Qwen3-32B  ──SFT (full data)──▶  Polyhymnia-32B-SFT-LoRA-v1.0  (this repository)

Intended use

Multi-turn roleplay and companion-style conversation.
Tool-calling assistants that emit a send_message call.
Research on persona consistency and conversational style adaptation.

Out of scope: factual question answering, medical / legal / financial advice, high-stakes decision making.

Training

Method: supervised fine-tuning (SFT) with LoRA.
Training data: 129,482 internal multi-turn conversations with tool-call annotations. The training data is not released with this adapter.
Maximum sequence length: 22,528 tokens.
Epochs: 1.
Effective batch size: 16 (per-device batch 1 × gradient accumulation 16).
Learning rate: 1e-4.
LoRA: r = 32, alpha = 32, dropout = 0.

Evaluation

Automatic evaluation on an internal held-out split (token-level eval loss):

Table with columns: Step, Eval loss, Note
Step	Eval loss	Note
4,200	0.6023	Best value recorded in the training log
4,300	0.6029
4,500	0.6058

Usage

PEFT / Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-32B"adapter_id = "PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0" tokenizer = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(    base_id, torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, adapter_id)model.eval()

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-32B"adapter_id = "PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0" tokenizer = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(    base_id, torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, adapter_id)model.eval()

vLLM

bash

vllm serve Qwen/Qwen3-32B \  --enable-lora --max-lora-rank 32 \  --lora-modules polyhymnia=PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0 \  --enable-auto-tool-choice --tool-call-parser hermes

vllm serve Qwen/Qwen3-32B \  --enable-lora --max-lora-rank 32 \  --lora-modules polyhymnia=PresenceAI/Polyhymnia-32B-SFT-LoRA-v1.0 \  --enable-auto-tool-choice --tool-call-parser hermes

Tool-call format used during training:

text

<tool_call>{"name": "send_message", "arguments": {"text": "..."}}</tool_call>

Use the tokenizer and chat_template.jinja included in this repository.

Limitations

Trained for roleplay/companion conversation; may be inappropriate outside that domain.
May reproduce patterns or content from its training data. The training data is not released; deploy with privacy and safety filters.
Only automatic eval loss is reported. No factual-accuracy, safety, or human-quality claim is made.
The base model's limitations and license terms continue to apply.

License

The adapter is released under the Apache License 2.0. The base model Qwen/Qwen3-32B is also Apache-2.0. See LICENSE and NOTICE.

Citation

bibtex

@misc{presenceai_polyhymnia_32b_sft_lora_v1,  title  = {PresenceAI Polyhymnia-32B-SFT-LoRA v1.0: roleplay and tool-calling chat},  author = {PresenceAI},  year   = {2026},  note   = {LoRA adapter, Apache-2.0}}

@misc{presenceai_polyhymnia_32b_sft_lora_v1,  title  = {PresenceAI Polyhymnia-32B-SFT-LoRA v1.0: roleplay and tool-calling chat},  author = {PresenceAI},  year   = {2026},  note   = {LoRA adapter, Apache-2.0}}