Stage-2 LoRA adapter from a two-stage SFT recipe on Qwen 3.6 27B. This adapter teaches Hermes-format tool calling (<tool_call>{...JSON...}</tool_call>) on top of a reasoning-tuned base. It is not standalone — apply it to the stage-1 reasoning base below.

Lineage: Qwen/Qwen3.6-27B → stage-1 reasoning LoRA → merge → this LoRA (tool-calling SFT).

Where this fits in the release

Artifact	Repo
Stage-1 BF16 merged (base for this LoRA)	`samscrack/Qwen3.6-27B-Opus-CoT-Stage1`
Stage-2 LoRA adapter (this repo)	`samscrack/Qwen3.6-27B-Hermes-S2-LoRA`
Stage-1 + Stage-2 merged + FP8 quantized (final release)	`samscrack/Qwen3.6-27B-Opus-CoT-S1-Hermes-S2-SFT`

If you just want the production model, use the FP8 repo above. This repo is for users who want to:

mix-and-match: apply the tool-calling stage to a different base, or
inspect / re-train / re-quantize stage 2 separately, or
merge yourself with different precision (BF16 / int8 / int4) than the official FP8.

Quick start (PEFT)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id    = "samscrack/Qwen3.6-27B-Opus-CoT-Stage1"
adapter_id = "samscrack/Qwen3.6-27B-Hermes-S2-LoRA"

tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

# Optional: bake the LoRA into the base for serving
# merged = model.merge_and_unload()
# merged.save_pretrained("./Qwopus3.6-27B-stage2-bf16")

Tool-call format (Hermes)

Tool calls are rendered as:

<tool_call>{"name": "function_name", "arguments": {"k": "v"}}</tool_call>

This matches the Hermes / Qwen3 family convention and is supported natively by vLLM (--tool-call-parser hermes) and SGLang. Servers that expect DeepSeek-native tool tokens, OpenAI function-call JSON, or Llama 3 <|python_tag|> form will need an adapter.

Training — Stage 2 only


Method	Supervised fine-tuning, LoRA via Unsloth + TRL `SFTTrainer`
Base	`samscrack/Qwen3.6-27B-Opus-CoT-Stage1` (BF16 merged stage-1)
LoRA	r=16, α=16, dropout=0, targets: `q_proj, k_proj, v_proj, o_proj, out_proj, gate_proj, up_proj, down_proj`
Optimizer / LR	AdamW, 5e-5, cosine schedule
Schedule	1 epoch, batch 2 × grad_accum 8 → effective batch 32, ctx 16384
Steps / final loss	74 / 0.334
Wall clock	~1 h 45 m on 2× RTX PRO 6000 Blackwell, DDP via `torchrun --standalone --nproc-per-node=2`

Dataset

Dataset	Rows	Notes
`DJLougen/hermes-agent-traces-filtered`	3,679	Pre-cleaned subset of `lambda/hermes-agent-reasoning-traces`; tool calls are valid JSON

Source corpus: lambda/hermes-agent-reasoning-traces (Kimi + GLM-5.1 configs, ~14.7k raw rows; ~13% are dropped by the trainer's JSON validator).

A strip_tool_response_wrappers helper in the training pipeline avoids the chat template re-wrapping <tool_response> tokens already present in the dataset.

Software

PyTorch 2.8.0+cu128, Transformers 5.2.0, TRL 0.22.2, PEFT 0.19.1, Unsloth 2026.4.7, datasets 4.3.0.

Limitations

Useless without the right base — apply specifically to samscrack/Qwen3.6-27B-Opus-CoT-Stage1 (or, with caveats, to a stage-1 you trained yourself with the same r=64 reasoning LoRA recipe).
The adapter targets all eight linear projection groups; merging requires the standard PEFT merge_and_unload path.
No RLHF / DPO step — supervised only.
74 steps over 3.7k rows is a short run by design (the source dataset is small after filtering); the resulting tool-call format adherence is high but coverage of rare tool patterns is modest.

Acknowledgements

The two-stage recipe is adapted from Jackrong's Jackrong-llm-finetuning-guide.

Dataset authors: DJLougen (filtered set), lambda (upstream traces).

License

Apache 2.0, inherited from the Qwen 3.6 base. Dataset licenses apply to derived behavior — see each dataset card.