CommentLens — fine-tuned Qwen2.5-7B comment analyzer
Fine-tune of Qwen/Qwen2.5-7B-Instruct that classifies a single YouTube comment
across eight dimensions and returns them as JSON. It powers
CommentLens, an app that turns a
thousand comments into a clear picture of what an audience is actually saying.
Published as a LoRA adapter (QLoRA, r=64) on top of Qwen/Qwen2.5-7B-Instruct. Apply it with vLLM's LoRA support or PEFT.
Output schema
Given one comment, the model returns:
| field | description |
|---|---|
valence |
positive / negative / neutral / mixed |
intent |
appreciation, question, request, personal_experience, humor, ... |
emotion |
gratitude, excitement, hope, confusion, sarcasm, ... |
stance |
trust, skepticism, certainty, uncertainty, curiosity |
target |
creator, video_content, topic, self_life, ... |
orientation |
passive_reaction, wants_more_info, ready_to_act, ... |
persona |
beginner, expert, aspirational, critic_troll, loyal_fan |
feedback |
clarity, depth, accuracy, pacing, relevance |
Prompt format (ChatML)
<|im_start|>system
You are a YouTube comment analyzer. Analyze the given comment and classify it
across multiple dimensions. Return a JSON object with these fields: valence,
intent, emotion, stance, target, orientation, persona, feedback<|im_end|>
<|im_start|>user
Analyze this YouTube comment:
{comment}<|im_end|>
<|im_start|>assistant
Use greedy decoding (temperature=0) for stable JSON.
Quick start (vLLM)
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest
from huggingface_hub import snapshot_download
adapter = snapshot_download("vellumy/commentlens-qwen")
llm = LLM(model="Qwen/Qwen2.5-7B-Instruct", enable_lora=True, max_lora_rank=64,
dtype="bfloat16", max_model_len=2048)
# ... build the ChatML prompt above, then:
out = llm.generate([prompt], SamplingParams(temperature=0, max_tokens=128,
stop=["<|im_end|>"]),
lora_request=LoRARequest("commentlens", 1, adapter))
print(out[0].outputs[0].text)
In CommentLens, set MODEL_PROVIDER=local and HF_MODEL_ID=vellumy/commentlens-qwen.