File size: 2,582 Bytes
d65c969 d638efc d65c969 feac4b1 d65c969 4aa1ea7 d65c969 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen
- qwen3.5
- fable-5
- claude-opus
- distillation
- merge
- bf16
base_model:
- Qwen/Qwen3.5-9B
- empero-ai/Qwable-9B-Claude-Fable-5
- Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled
- Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2
---
# FableOpus 9B Delta bf16
Opus-forward delta merge anchored to Qwen3.5-9B, keeping Fable tool-use behavior while biasing toward the high-performance Opus reasoning family.
This is a **bf16 safetensors merge** in the Qwen3.5-9B family. It combines the agentic/tool-use flavor of Fable 5 distillation with Claude Opus reasoning distilled checkpoints.
## Recipe
- Base anchor: `Qwen/Qwen3.5-9B`
- Merge method: `delta_linear`
- Output dtype: `bfloat16`
Weights:
- `empero-ai/Qwable-9B-Claude-Fable-5`: 0.38
- `Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled`: 0.42
- `Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2`: 0.2
The local `mergekit`/`transformers` stack did not yet recognize the new `qwen3_5` model type, so the merge was performed directly tensor-by-tensor over compatible safetensors checkpoints. Non-floating tensors are copied from the Fable/Qwable checkpoint; floating tensors are emitted as bf16.
## Source Signals
- Fable source: `empero-ai/Qwable-9B-Claude-Fable-5`, derived from Fable 5 traces.
- Opus source: `Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled`, a high-download Opus reasoning distilled checkpoint.
- Opus v2 source: `Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2`.
## Intended Use
General chat, code assistance, tool-use style prompting, and reasoning-heavy experiments. Evaluate before production use. This model inherits limitations and licensing/provenance constraints from its source checkpoints and datasets.
## Quick Start
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "interpolators/FableOpus-9B-Delta"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
messages = [{"role": "user", "content": "Write a concise plan for building a small agentic coding benchmark."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tok.decode(out[0], skip_special_tokens=True))
```
|