Downloads · 30 days
124
36% of all-time downloads
AlisherAmirbek/Qwen3.6-35B-A3B-heretic-lora
Qwen3.6-35B-A3B-heretic-lora is a text generation model from AlisherAmirbek. Use it when you need the model to write or continue text. It is set up for peft.
Stage 2 LoRA for tool selection, search-filter and geolocation extraction, and profile-signal extraction. Trained fresh from the base model, without a dialog adapter or an earlier specialized adapter as initialization.
Downloads · 30 days
124
36% of all-time downloads
All-time downloads
343
Public
Repo size
83.9 GB
Likes
0
Public
Click a slice to open those files.
.safetensors3.7 GB · 99%
From the Hugging Face model README
Stage 2 LoRA for tool selection, search-filter and geolocation extraction, and profile-signal extraction. Trained fresh from the base model, without a dialog adapter or an earlier specialized adapter as initialization.
Compared with release 80a89670, the training data is the same, but every
target is serialized more compactly:
"erotic-massage-parlors" for
HOT.com search, a tool name such as "refusal", or
{"route":"internal_memory","query":"..."} for routes that take a query.
Several routes per turn are still allowed. category_group is no longer
emitted; serving code derives it from the category name.country, city, state and neighborhood, for
example {"country":"us","city":"miami","state":"fl"}.entity_type
wrapper; no filters is {}.Serving code must parse these formats: this release does not emit XML root tool
calls. The matching code is in hot-llm merge request 196
(feat/compact-structured-outputs). The tokenizer, chat template and adapter
configuration are byte-identical to 80a89670.
The base tokenizer vocabulary of 248,077 tokens is retained; this release has no legacy agent-context tokens or trained embedding rows.
stage2-compact-outputs-20261109T185829Z, data and code from hot-llm
commit ef3d88bc (the run's checkout also had an unused --load-in-fp8
option; base weights were loaded in 4-bit).dbfd9eb0cdc7c33fc970b06429f6e043b6851190.80a89670.These weights have already been converted by
scripts/training/convert_unsloth_moe_lora_for_vllm.py: 80 expert LoRA-B tensors,
256 experts, rank 16. Logical-layout equivalence, unchanged non-target tensors,
and tensor-key preservation were verified. Do not apply that conversion again
to this export. See vllm_conversion.json for checksums and conversion details.
Load this export as a 3D MoE LoRA (is_3d_lora_weight=true). The full offline
evaluation used vLLM 0.26.0 with the supplied tokenizer and native tool-call chat
template.
All 4,667 evaluation outputs finished normally without generation truncation.
Strict target matching: 4,139 / 4,667 (88.69%), against 4,123 / 4,667 for
80a89670 on the same held-out set.
| Task | This release | 80a89670 |
|---|---|---|
| Root tool selection | 1,642 / 2,106 (78.0%) | 1,631 / 2,106 (77.4%) |
| Search filters | 730 / 769 (94.9%) | 736 / 769 (95.7%) |
| Geolocation | 846 / 862 (98.1%) | 845 / 862 (98.0%) |
| Profile signals | 921 / 930 (99.0%) | 911 / 930 (98.0%) |
Root tool choice is 2,074 / 2,106, with 275 / 276 refusals and one false refusal; 432 of the 464 strict root misses pick the right tools and differ only in arguments such as query wording. Of the 310 held-out turns that need a knowledge tool, 293 get the right tool set, and the two knowledge tools are confused with each other 7 times.
Mean generated tokens per output fell from 32.2 to 9.0 for root tool selection (median 36 to 5), 23.6 to 10.9 for search filters, 18.0 to 12.2 for geolocation, and 43.6 to 32.5 for profile signals.
The 14 search-filter samples that 80a89670 matched and this release misses
were reviewed: each has a wrong, missing or extra value, not a format error. The
other mismatches were not manually reviewed. These are offline benchmark
results; live application behavior has not been evaluated for this release.