CARDS-Wind-Qwen3.6-27B-FP8

Joint climate-discourse classifier — a single Qwen3.6-27B backbone fine-tuned on CARDS + Wind concatenated, then FP8-dynamic quantized. One model handles both tasks:

• CARDS — classification of climate-contrarian claims under the Coan et al. (2025) hierarchical taxonomy.
• Wind — three-level wind-energy opposition classification (detection / frames / claims).

The model picks the right task from the system prompt: pass the CARDS system prompt for CARDS-style output, the Wind system prompt for Wind-style output. Same weights, same chat template.

This is the FP8 deployment variant — ~27 GB on disk, fits on a single A100/H100/H200.

Why a joint model

• One checkpoint, two tasks. Saves disk + VRAM + ops complexity if you serve both classifiers in production.
• Same recipe, additive cost. Trained with cards/ft/train.py --joint — Unsloth + LoRA r=16 α=16, 3 epochs, lr=2e-4 — identical to the cards FT models, with the wind dataset added to the training mix.
• Wind paper is forthcoming. This model accompanies a forthcoming C3DS paper on wind-energy opposition discourse — research-preview status until the paper is out.

Results (Wind test set)

Evaluated on the wind-opposition test set (773 rows, 436 opposition-positive). Compared with the wind-only sibling, the BF16 joint model, and frontier APIs:

Detection (binary)

Metric	Windy-27B FP8	CARDS-Wind-27B (BF16)	CARDS-Wind-27B (FP8 — this model)	Claude Opus 4.7	GPT-5.5
Precision	0.877	0.863	0.866	0.896	0.927
Recall	0.920	0.911	0.917	0.890	0.846
F1	0.898	0.886	0.891	0.893	0.885

Samples F1 (multi-label frame / claim accuracy)

View	Windy-27B FP8	CARDS-Wind-27B (BF16)	CARDS-Wind-27B (FP8)	Claude Opus 4.7	GPT-5.5
Frames — all rows	0.787	0.770	0.772	0.791	0.792
Frames — opposition only	0.751	0.736	0.739	0.734	0.697
Claims — all rows	0.755	0.733	0.738	0.754	0.745
Claims — opposition only	0.694	0.668	0.677	0.667	0.614

• Joint training costs ~0.007 detection F1 vs the wind-only Windy-27B-FP8 (0.891 vs 0.898) — small but real.
• Still ties or beats Claude Opus 4.7 on detection F1 (0.891 vs 0.893) and beats it on frames-opposition-only and claims-opposition-only samples F1.
• FP8 ≈ BF16 on every metric — quantization is effectively free here.
• Zero parse failures on 773 test items.

CARDS test set

CARDS-side metrics for the joint model are not separately reported in this release — the cards-only sibling C3DS/CARDS-Qwen3.6-27B is the canonical reference for CARDS test-set numbers (samples F1 = 0.893 at L1, ties Opus 4.6). The joint model uses an identical CARDS training setup; expect comparable performance, but treat as approximate until reproduced.

Usage

Routing between the two tasks

The model picks its behaviour from the system prompt. Both prompts are bundled in this repo:

• cards_prompts.json — slim_system_instruction (CARDS) + cot_trigger
• wind_prompts.json — slim_system_instruction (Wind)

With vLLM

vllm serve C3DS/CARDS-Wind-Qwen3.6-27B-FP8 \
  --port 8000 \
  --max-model-len 4096 \
  --enable-prefix-caching \
  --kv-cache-dtype fp8 \
  --served-model-name CARDS-Wind-Qwen3.6-27B

import json
from huggingface_hub import hf_hub_download
from openai import OpenAI

cards = json.load(open(hf_hub_download("C3DS/CARDS-Wind-Qwen3.6-27B-FP8", "cards_prompts.json")))
wind  = json.load(open(hf_hub_download("C3DS/CARDS-Wind-Qwen3.6-27B-FP8", "wind_prompts.json")))

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

def classify_cards(text):
    resp = client.chat.completions.create(
        model="CARDS-Wind-Qwen3.6-27B",
        messages=[
            {"role": "system", "content": cards["slim_system_instruction"]},
            {"role": "user",   "content": f"### Text:\n{text}\n\n{cards['cot_trigger']}"},
        ],
        temperature=0, max_tokens=4000,
    )
    return resp.choices[0].message.content

def classify_wind(text):
    resp = client.chat.completions.create(
        model="CARDS-Wind-Qwen3.6-27B",
        messages=[
            {"role": "system", "content": wind["slim_system_instruction"]},
            {"role": "user",   "content": text},
        ],
        temperature=0, max_tokens=4000,
    )
    return resp.choices[0].message.content

Both modes produce a <think>…</think> reasoning trace followed by a YAML block. CARDS output is categories: [...]; Wind output is opposition_detected, frames, claims.

Multimodal — image + text

The base Qwen3.6-27B supports image inputs via the OpenAI-compatible image_url content part, and this fine-tune preserves that capability for both tasks. Switch tasks by switching the system prompt — CARDS prompt for CARDS-style output, Wind prompt for Wind-style output — and pass an image (with or without caption text) alongside.

Serve vLLM with multimodal flags enabled:

vllm serve C3DS/CARDS-Wind-Qwen3.6-27B-FP8 \
  --port 8000 \
  --max-model-len 8192 \
  --trust-remote-code \
  --limit-mm-per-prompt image=4 \
  --enable-prefix-caching \
  --kv-cache-dtype fp8 \
  --served-model-name CARDS-Wind-Qwen3.6-27B

import base64, json, mimetypes
from pathlib import Path
from huggingface_hub import hf_hub_download
from openai import OpenAI

cards = json.load(open(hf_hub_download("C3DS/CARDS-Wind-Qwen3.6-27B-FP8", "cards_prompts.json")))
wind  = json.load(open(hf_hub_download("C3DS/CARDS-Wind-Qwen3.6-27B-FP8", "wind_prompts.json")))

def image_part(path):
    p = Path(path)
    mime = mimetypes.guess_type(p)[0] or "image/png"
    b64 = base64.b64encode(p.read_bytes()).decode()
    return {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}}

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

# Wind classification on an image:
resp = client.chat.completions.create(
    model="CARDS-Wind-Qwen3.6-27B",
    messages=[
        {"role": "system", "content": wind["slim_system_instruction"]},
        {"role": "user", "content": [
            {"type": "text", "text": "Read the image and any caption below; classify the wind-opposition framing depicted."},
            image_part("screenshot.png"),
            {"type": "text", "text": "### Caption:\n<optional caption>"},
        ]},
    ],
    temperature=0,
    max_tokens=4000,
)
print(resp.choices[0].message.content)

Training & Quantization

Joint fine-tuning

• Base model: Qwen/Qwen3.6-27B
• Method: LoRA (rank 16, α 16, dropout 0) on q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, then merged into base weights
• Datasets: C3DS/cards_sft_dataset (CARDS RECoT messages) + the wind RECoT messages corpus (forthcoming release)
• Mix: datasets concatenated row-wise — no balancing or task-token; the model learns to pick its task from the system prompt
• Framework: Unsloth + TRL SFTTrainer, invoked via cards/ft/train.py --joint
• Hyperparameters: 3 epochs, per_device_train_batch_size=1, gradient_accumulation_steps=8, lr=2e-4, cosine schedule, 10 warmup steps, max_seq_length=8192, adamw_8bit, bf16

FP8 quantization

• Scheme: fp8_e4m3 dynamic per-channel quantization (weights only). Activations stay in BF16; no calibration data required.
• Targets: linear layers in transformer blocks; lm_head left in BF16.
• Tool: llmcompressor with QuantizationModifier(targets="Linear", scheme="FP8_DYNAMIC") applied to the merged BF16 joint checkpoint.

Limitations

• Forthcoming Wind paper. Wind-side methodology, codebook, and dataset details are pending publication.
• CARDS-side metrics not separately re-evaluated for the joint model — refer to the cards-only sibling for the canonical CARDS test-set numbers.
• Joint training trade-off. Detection F1 on Wind is marginally lower than the wind-only Windy-Qwen3.5-27B-FP8 (0.891 vs 0.898). Use the dedicated wind model if absolute wind detection F1 is the priority; use this one if you need both tasks from a single backbone.
• Thinking tokens. Training used enable_thinking=True. Parse output after </think> or disable thinking at inference.

Related models

• C3DS/CARDS-Qwen3.6-27B — CARDS-only Qwen3.6-27B sibling (same backbone, CARDS-only training).
• C3DS/Windy-Qwen3.5-27B-FP8 — Wind-only 27B FP8 sibling (Qwen3.5 backbone).

Citation

For the CARDS side of this model, please cite:

@article{coan2025cards,
  title   = {Large language model reveals an increase in climate contrarian speech in the United States Congress},
  author  = {Coan, Travis G. and Malla, Ranadheer and Nanko, Mirjam O. and Kattrup, William and Roberts, J. Timmons and Cook, John and Boussalis, Constantine},
  journal = {Communications Sustainability},
  volume  = {1},
  pages   = {37},
  year    = {2025},
  doi     = {10.1038/s44458-025-00029-z}
}

A wind-side citation will be added when the corresponding paper is published.

License

Apache 2.0, inherited from Qwen3.6-27B.