vheart-affect-v8

A LoRA adapter that turns a Qwen2.5-1.5B-Instruct base into a fine-tuned affect source for the feltstate felt-state library.

This is the small / fast reference. For finer mixed-feeling resolution and the 50-label vocab, see kaishuiji/vheart-affect-v9.

What it does

Reads a short dialogue turn (assistant + user history) and emits an AffectDelta-shaped JSON:

{
  "valence": -0.31,
  "arousal": 0.52,
  "labels": ["frustrated", "curious"],
  "confidence": 0.78,
  "monologue": "想搞懂但卡住了，又想再试一次",
  "mixed_blend": {
    "primary": "frustrated",
    "secondary": "curious",
    "weights": [0.6, 0.4]
  }
}

This is the LLMSource slot in feltstate's pipeline — measurement, not generation. The reply model reads the resulting felt-state block and decides for itself how to act.

Why it exists

feltstate.PHILOSOPHY §1 says: "the real signal comes from an LLMSource (a separate model call whose only job is to measure) or a classifier you fine-tune."

This is the second half of that sentence — an actual fine-tuned classifier.

Training data — not released

~750 SFT samples, hand-curated Chinese-first dialogue snippets with labeled affect outcomes. Data not released to protect contributor privacy. The schema and label vocabulary are documented so others can build their own corpus.

Label vocab — 24 labels (v8)

Anchor table (valence, arousal) drawn from internal rater calibration:

label	v	a	notes
focused	+0.58	0.45	mid-arousal — quiet focus, not hype
frustrated	-0.48	0.37	low arousal — stuck, not exploding
satisfied	+0.76	0.59
tired	-0.67	0.38	low-v low-a — depleted
exploring	+0.30	0.53	curious-leaning, slightly positive
...			(24 total)

Inter-rater agreement on these labels: r ≈ 0.74 on calibration set. (See v9 for the 50-label expansion.)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
mdl = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
mdl = PeftModel.from_pretrained(mdl, "kaishuiji/vheart-affect-v8")
mdl.eval()

prompt = "..."  # see prompt template below
out = mdl.generate(tok(prompt, return_tensors="pt").input_ids.to(mdl.device),
                   max_new_tokens=256, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))

With feltstate

from feltstate import Engine
from feltstate.sources import VheartSource  # see feltstate/sources/vheart.py

eng = Engine(source=VheartSource("kaishuiji/vheart-affect-v8"))
eng.observe("今晚跑通了三周的实验。")
print(eng.state.mood)

Prompt template

Input is a chat-style turn; assistant response is the JSON above. System prompt (translated; production uses Chinese):

You measure the affect of a character in this dialogue. You do not generate replies; you output JSON only. Schema fields: valence (-1..1), arousal (0..1), labels (subset of vocab), confidence (0..1), monologue (short Chinese sentence — what the character feels internally, not says aloud), mixed_blend (primary + secondary label with weights). Same label, different context → different point. Do not collapse to templates.

Full system prompt available in the prompt_template.txt file in this repo.

Limitations

• Chinese-first. English inputs work but trained mostly on Chinese.
• Single-character measurement. Designed for a felt state of one ongoing companion, not third-party emotion analysis at scale.
• Schema bound to feltstate. Labels are the v8 24-vocab. If you need a different vocab, fine-tune your own — this is a reference, not a hammer.
• 750 samples is small. v9 expanded data; if you have GPU budget, use v9.

Citation

@software{feltstate_vheart,
  author = {morephine},
  title = {feltstate / vheart-affect-v8},
  url = {https://huggingface.co/kaishuiji/vheart-affect-v8},
  year = {2026}
}

License

Apache 2.0 (matches Qwen2.5 base).

Acknowledgments

• Qwen team for the base model
• feltstate philosophy authors (same hands)