Chimera-14B — DeepSeek-R1-Distill-Qwen-14B Agent LoRA

A dual-capability LoRA adapter for huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2 that adds tool calling and extended reasoning to the same weights. The model reasons inside <think> blocks and emits tool-call JSON in the same response.

Two heads, one beast: one for thought, one for action.

• Base model: huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2
• Adapter type: LoRA (PEFT), r=16, alpha=16, dropout=0, bias=none
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trainable params: 68,812,800 (0.46% of 14.8B)
• Chat template: qwen-2.5
• License: MIT (inherited from the DeepSeek-R1-Distill-Qwen-14B lineage)

Why this model exists

DeepSeek-R1-Distill-Qwen-14B is already remarkable at one thing: packing genuinely impressive reasoning into a small parameter count. The distillation process compressed R1's deliberative thinking down to a 14B footprint, and the result thinks in ways that feel far larger than its size.

Chimera-14B takes that and makes it useful.

A model that only reasons is a model that thinks beautifully and then stops. The missing piece is the ability to act on those thoughts — to call tools, fetch data, and close the loop between deliberation and execution. So this adapter stacks two capabilities on top of the distilled reasoning base:

1. Tool calling — the model learns to emit structured tool-call JSON, turning its reasoning into concrete actions.
2. Extended reasoning — further SFT on top deepens and lengthens the <think> chains it can sustain.

The result is a 14B model that reasons like a much larger one and acts on what it reasons about — a genuinely impressive amount of reasoning for a model this small, with the tool-calling to make it do something.

What it does

Trained in two sequential stages on the same LoRA weights:

1. Stage 1 — tool calling on DJLougen/hermes-agent-traces-filtered: teaches the model to emit structured tool-call JSON.
2. Stage 2 — reasoning on open-r1/Mixture-of-Thoughts (config all, 2,425 train / 50 eval examples, filtered to <1,900 tokens): teaches the model to reason inside <think> blocks.

Verified at inference: the model produces a <think> reasoning block and a tool-call JSON payload in the same response — no separate passes, no scaffolding.

Quickstart — PEFT (transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2"
adapter_id = "Flexingmeow/Chimera-14B"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather for a city",
            "parameters": {
                "type": "object",
                "properties": {"city": {"type": "string"}},
                "required": ["city"],
            },
        },
    }
]

messages = [
    {
        "role": "user",
        "content": "What's the weather in Osaka right now? Use the weather tool.",
    }
]
text = tokenizer.apply_chat_template(
    messages, tools=tools, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.7,
        do_sample=True,
    )

print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Important: pass tools=... to apply_chat_template. The model is trained to call the tools it is shown; without tool definitions in the prompt it will answer conversationally (or refuse-style) instead of emitting a tool call.

Expected output shape

Verified at inference, the model emits a <think> reasoning block followed by a markdown-fenced JSON block — not <tool_call> tags:

<think>
The user is asking about weather in Osaka, so I should call the weather tool with the city parameter set to "Osaka".
</think>

```json
{"name": "get_weather", "arguments": {"city": "Osaka"}}
```

The JSON content is a valid tool call, but the wrapper is a markdown code fence. Write your parser to extract the JSON from inside the ```json block rather than matching <tool_call> tags — the model does not emit those.

Quickstart — llama.cpp (GGUF LoRA)

This repo bundles everything you need — the base model GGUF (DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.Q4_K_M.gguf) plus the adapter (agent-lora-f16.gguf, 137.6 MB, f16, llama.cpp LoRA format). Apply the adapter on top of the base:

# clone the repo (or download the two .gguf files from the Files tab)
git lfs install && git clone https://huggingface.co/Flexingmeow/Chimera-14B

llama-cli \
  -m DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.Q4_K_M.gguf \
  --lora agent-lora-f16.gguf \
  --jinja \
  -ngl 99 \
  -c 4096 \
  --temp 0.7 \
  -p "What's the weather in Osaka right now? Use the weather tool."

• --lora agent-lora-f16.gguf applies the adapter (the short flag is -l, but --lora is unambiguous).
• --jinja is required — it enables the repo's chat template (qwen-2.5), which is what makes tool calling work.
• -ngl 99 offloads all layers to GPU; drop it if you're running CPU-only.

Use --lora-scaled agent-lora-f16.gguf <scale> to tune adapter strength (default scale is 1.0).

Training details

Two sequential SFT stages on the same adapter weights, trained with Unsloth on a single Tesla T4 (Kaggle), ~5.5 hours total for stage 2.

Stage 1 — tool calling

• Dataset: DJLougen/hermes-agent-traces-filtered
• Output: qwen-14b-tool-calling-lora-final

Stage 2 — reasoning

• Dataset: open-r1/Mixture-of-Thoughts (config all)
• Splits: 2,425 train / 50 eval, filtered to <1,900 tokens per example
• Epochs / steps: 1 epoch, 304 steps
• Output: qwen-14b-agent-lora-final

Hyperparameters

Parameter	Value
LoRA r	16
LoRA alpha	16
LoRA dropout	0
Bias	none
Learning rate	5e-5
LR schedule	cosine
Warmup steps	20
Batch size	1 × grad accum 8 (effective 8)
Optimizer	adamw_8bit
Weight decay	0.01
Precision	fp16
Max sequence length	2048
Packing	True

Evaluation

Eval loss during stage 2:

Step	Eval loss
50	0.9476
300	0.8695

Files

File	Description
adapter_model.safetensors	LoRA weights (PEFT format, 275 MB)
adapter_config.json	PEFT adapter config
agent-lora-f16.gguf	LoRA in llama.cpp GGUF format (f16, 137.6 MB)
DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.Q4_K_M.gguf	Base model GGUF (Q4_K_M, ~9 GB) — included so the LoRA works out of the box with llama.cpp
chat_template.jinja	Qwen-2.5 chat template
tokenizer.json, tokenizer_config.json	Matching tokenizer files

Limitations

• Trained on an abliterated base model; inherit the usual caveats about uncensored weights.
• Tool-calling quality is strongest for JSON-schema-style tools; exotic argument formats may need few-shot prompting.
• Tool-call output is wrapped in a markdown ```json fence, not <tool_call> tags — parsers must extract from the fence.
• 14B at fp16 needs ~28 GB VRAM for the full base + adapter; use 4-bit quantization of the base for smaller GPUs.