qwen35-9b-kuairand-recexplain-lora API & Inference Endpoint

Intended Use

The adapter is intended for research and demonstration in:

User interest summarization.
Candidate item recommendation explanation.
Preference-aware reasoning over user histories, item metadata, and Semantic ID tokens.

It is not intended as a general-purpose assistant.

Training Summary

The adapter was trained with:

Teacher-enhanced SFT.
Warm-start DPO over hard/medium/random negative preference pairs.
Teacher-judged student-aware DPO over sampled student responses.

The final released adapter corresponds to the student-aware DPO checkpoint-600 selected by held-out generation and preference-ranking evaluation.

The derived SFT / DPO post-training data and generated Semantic ID assets are released separately at plumliu/kuairand-recexplain-sft-dpo-data. The RQ-VAE checkpoint used to construct Semantic IDs is released at plumliu/kuairand-semantic-id-rqvae.

Evaluation

Table with columns: Metric, SFT, Released Adapter, Gain
Metric	SFT	Released Adapter	Gain
Candidate item grounding	`94.5%`	`97.5%`	`+3.0 pp`
Preference seq accuracy	`65.8%`	`~99.8%`	`+34.0 pp`

Loading Example

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3.5-9B"
adapter = "plumliu/qwen35-9b-kuairand-recexplain-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)

Limitations

The adapter is domain-specific and optimized for recommendation explanation prompts.
The model repository does not include training data, teacher API outputs, item-level Semantic ID mappings, or the RQ-VAE checkpoint.
Output quality depends on prompt format and the availability of grounded item/user context.

Intended Use

The adapter is intended for research and demonstration in:

User interest summarization.
Candidate item recommendation explanation.
Preference-aware reasoning over user histories, item metadata, and Semantic ID tokens.

It is not intended as a general-purpose assistant.

Training Summary

The adapter was trained with:

Teacher-enhanced SFT.
Warm-start DPO over hard/medium/random negative preference pairs.
Teacher-judged student-aware DPO over sampled student responses.

The final released adapter corresponds to the student-aware DPO checkpoint-600 selected by held-out generation and preference-ranking evaluation.

Evaluation

Table with columns: Metric, SFT, Released Adapter, Gain
Metric	SFT	Released Adapter	Gain
Candidate item grounding	`94.5%`	`97.5%`	`+3.0 pp`
Preference seq accuracy	`65.8%`	`~99.8%`	`+34.0 pp`

Loading Example

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3.5-9B"
adapter = "plumliu/qwen35-9b-kuairand-recexplain-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)

Limitations

The adapter is domain-specific and optimized for recommendation explanation prompts.
The model repository does not include training data, teacher API outputs, item-level Semantic ID mappings, or the RQ-VAE checkpoint.
Output quality depends on prompt format and the availability of grounded item/user context.

qwen35-9b-kuairand-recexplain-lora

README

Intended Use

Training Summary

Evaluation

Loading Example

Limitations

Explore FriendliAI today

README

Intended Use

Training Summary

Evaluation

Loading Example

Limitations