HyperCLOVAX-SEED-Think-32B-heretic

HyperCLOVAX-SEED-Think-32B-heretic은 naver-hyperclovax/HyperCLOVAX-SEED-Think-32B를 기반으로, 사후(weight editing) 방식으로 과잉 거부(refusal) 성향을 완화하는 방향의 수정이 적용된 변형 모델입니다.

---

Model Summary (KO)

• Base model: naver-hyperclovax/HyperCLOVAX-SEED-Think-32B
• Format: BF16 (safetensors)
• Approach: targeted post-hoc weight editing
• Primary effect: refusal behavior calibration (reduce over-refusal)
• Observed drift: small (see KL metric below)

---

Model Summary (EN)

• Base model: naver-hyperclovax/HyperCLOVAX-SEED-Think-32B
• Weights: BF16 (safetensors)
• Method: targeted post-hoc weight editing
• Goal: reduce over-refusal on benign/borderline prompts while keeping the output distribution close to the base model

---

What’s Changed

This variant applies focused modifications around projection layers to shift refusal-related behavior.

Editing Parameters (as-run)

• direction_index = 42.77
• attn.o_proj.max_weight = 1.13
• attn.o_proj.max_weight_position = 67.44
• attn.o_proj.min_weight = 0.46
• attn.o_proj.min_weight_distance = 25.36
• mlp.down_proj.max_weight = 1.49
• mlp.down_proj.max_weight_position = 43.36
• mlp.down_proj.min_weight = 0.97
• mlp.down_proj.min_weight_distance = 26.08

Behavior / Drift Indicators

• Refusals: 10 / 100
• KL divergence: 0.1123

Note: these values depend on your prompt set and refusal criterion.
If you can disclose the evaluation slice (even roughly), add it in the section below.

---

Evaluation Notes

Refusal Definition (edit as needed)

A completion is counted as a refusal if it:

• explicitly denies compliance (e.g., “I can’t / I won’t”), and
• does not provide a meaningful safe alternative or partial completion.

Prompt Set

• prompt mix: [benign / borderline / policy-sensitive]
• sample size: 100
• source: [private/internal or 공개 가능하면 설명]

---

Intended Use

Recommended

• General chat
• Creative writing / brainstorming
• Everyday Q&A where over-refusal hurts usability
• Research on refusal behavior, steering, and drift tradeoffs

Not Recommended (without extra guardrails)

• Public-facing deployment without moderation/filters
• High-stakes domains (medical/legal/financial)
• Any use that requires strict compliance guarantees

---

Safety & Risks

Reducing refusals can increase the chance that the model responds in situations where the base model would refuse. For real deployments, consider:

• input filtering / output moderation
• rate limits & logging
• clear acceptable-use policy and enforcement

Known limitations:

• side effects may exist (tone shift, verbosity changes, occasional riskier completions)
• evaluation is not exhaustive; additional red-teaming is recommended

---

GGUF (llama.cpp) Inference

This repository also provides an F16 GGUF build under gguf/, intended for running with llama.cpp.

Run with llama-server (Thinking ON)

This command enables the model's "thinking" behavior via --chat-template-kwargs.

Linux / macOS

./llama-server \
  -m {PATH}/HyperCLOVAX-SEED-Think-32B-heretic2.f16.gguf \
  --host 0.0.0.0 --port 10000 \
  --jinja \
  --chat-template-kwargs '{"thinking":true,"enable_thinking":true}' \
  -cb -fa on

---

## How to Use

### Transformers (example)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "hostkimjang/HyperCLOVAX-SEED-Think-32B-heretic"  # <- your repo id

tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain KL divergence in simple terms."},
]

# If the tokenizer provides a chat template:
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.95,
    do_sample=True,
)
print(tok.decode(out[0], skip_special_tokens=True))