---
license: apache-2.0
base_model: Qwen/Qwen3.6-27B
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- qwen
- qwen3
- qwen3.6
- carnice
- hermes-agent
- agentic
- sft
- bf16
- merged
---

![Carnice V2 27B](figures/carnice-v2-27b-hero.png)

# Carnice-V2-27B for Hermes Agent

Carnice-V2-27B is a full merged BF16 SFT of `Qwen/Qwen3.6-27B` for Hermes-style agent traces. This repository contains the standalone merged model weights, not only a LoRA adapter.

## BF16 Transformers Loading Fix

The BF16 safetensors were republished with corrected `Qwen3_5ForConditionalGeneration` tensor prefixes. The original merge artifact accidentally serialized an extra Unsloth wrapper prefix, which caused direct HF Transformers loads to report the real weights as unexpected keys and initialize expected layers randomly. GGUF files were not affected because the GGUF conversion path normalized those prefixes.

## Benchmarks

![Carnice V2 benchmark card](figures/benchmark-card.png)

| Metric | Qwen3.6-27B base | Carnice SFT |
| --- | ---: | ---: |
| IFEval prompt strict, limit 20 | 85.0% | 90.0% |
| IFEval prompt loose, limit 20 | 85.0% | 90.0% |
| IFEval instruction strict, limit 20 | 90.0% | 93.3% |
| IFEval instruction loose, limit 20 | 90.0% | 93.3% |
| Held-out assistant-token eval loss | 0.607 | 0.414 |
| Held-out assistant-token eval perplexity | 1.835 | 1.513 |

The benchmark artifact bundle is included under `benchmarks/`. It contains the rendered graph, extracted `metrics.json`, benchmark scripts, and raw result files used to make the chart.

Scope note: the IFEval run is a short `limit=20` A/B smoke benchmark, not an official full leaderboard score. Held-out loss/perplexity is the exact assistant-only training-format validation metric from the SFT script. The raw BFCL two-case smoke files are included for auditability, but they are too small to use as a model-quality claim.

## Training

This checkpoint was produced from the recovered 8K split-window Carnice run:

| Item | Value |
| --- | --- |
| Base model | `Qwen/Qwen3.6-27B` |
| SFT framework | Unsloth/PEFT LoRA, then merged to BF16 safetensors |
| Loss mask | Assistant-token-only |
| Context/windowing | 8,192 token windows with 1,024 token overlap |
| Train rows before windowing | 3,473 |
| Train windows | 6,554 |
| Eval examples | 110 |
| Source mix | 1,508 Carnice rows, 1,015 DJLougen Hermes rows, 950 Lambda GLM-5.1 Hermes rows |

## Usage

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

model_id = "kai-os/carnice-v2-27b"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
```

This is intended for agentic Hermes-style use. Validate with your own agent harness before relying on it for production behavior.