Modelpublic

hemlang/Hemlock-Qwen3.6-27B

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes10downloads

Model Card

Hemlock-Qwen3.6-27B

Vorarbeiter with the Schierling adapter merged in — an operator model that also writes Hemlock.

The combination exists because no code agent understands Hemlock. Vorarbeiter's delegation policy hands engineering work to a coding agent, which is right for every language that agent knows and wrong for this one: if the operator does not write Hemlock itself, nothing does.

What each half contributes

Measured on the TIES base, the Schierling adapter takes hembench from 10.1% to 65.0% (4/38 → 26/38 tasks), where every generated program is executed by the real interpreter and its stdout compared against expected output:

level		TIES	+ Schierling
L1	syntax	1/9	8/9
L2	stdlib	0/5	1/5
L3	algorithms	0/7	5/7
L4	systems	0/7	5/7
L5	translation	0/5	3/5
L6	debugging	3/5	4/5
overall	weighted	10.1%	65.0%

22 tasks fixed, 0 broken. Generated code that fails to execute drops from 30 tasks to 8.

Vorarbeiter contributes the operator behaviour: delegation of non-Hemlock engineering work to a code agent, and factual answers on topics the base sanitises. See its card for those numbers.

Benchmarks

Measured on the merged weights. Tool use is the 47-case egirl bench at temperature 0; Hemlock is hembench, where every generated program is executed by the real interpreter and its stdout compared against expected output.

	TIES	Vorarbeiter	Hemlock	Wichtel
tool use	35/47	41/47	—	47/47
delegation	0/10	7/10	—	10/10
hembench	10.1%	—	—	67.6%

The concern flagged when this model was uploaded — that 5,562 rows of tool-free code generation might teach the operator to write code instead of delegating — did not materialise. Adding the Hemlock adapter to a base with 0/10 delegation left it at 1/10 and lifted total tool use 35/47 → 37/47; nothing was lost. Adding the egirl tool-use adapter on top then took delegation to 10/10, which is Wichtel.

If you want the operator model, use Wichtel — same lineage plus the tool-use adapter, and it is the configuration that was benchmarked end to end. This model is the useful intermediate: the Hemlock capability without the egirl-specific tool schema baked in, which is what you want if your agent uses different tools.

MTP head

peft merge_and_unload() silently drops the 15 mtp.* tensors and current llama.cpp then refuses to load the model. They are grafted back from the pre-merge base and verified bit-identical. This architecture has 1,199 tensors (850 language, 333 vision, 15 MTP, 1 lm_head).

bash

llama-server -m Hemlock-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 4

Limitations

—stdlib is the weak spot (1/5 on the adapter's own evaluation). Check API calls against docs.
—Merge behaviour unbenchmarked at upload time — see above.
—38 hembench tasks is a small sample; per-level scores carry wide error bars.
—Everything Vorarbeiter carries: single-turn tool use, delegation bound to egirl's tool schema, and a persistent Chinese-language gap on the censorship axis.

Lineage

Qwen/Qwen3.6-27B
  └─ huihui-ai/Huihui-Qwen3.6-27B-abliterated
      └─ nbeerbower/Qwen3.6-27B-TIES
          └─ + Bubba-3ep (GreatFirewall ORPO)
              └─ + delegation LoRA          = Vorarbeiter
                  └─ + Schierling (Hemlock SFT)   ← this model