465e74f49e
fix: patch is_mllm_model for Qwen3.6 text-only model
...
AutoRound misidentifies Qwen3_5ForConditionalGeneration as a VLM
and tries to load a vision processor. Patch to force LLM mode.
2026-06-01 04:26:15 +02:00
367ed705ab
fix: convert chat messages to text for AutoRound calibration
2026-06-01 04:23:28 +02:00
4edaeeb21b
switch quantization from llm-compressor to AutoRound
...
llm-compressor pins transformers<=4.57.6, can't load Qwen3.6.
AutoRound (Intel) works with transformers 5.x and is already
installed as an llmcompressor dependency. Produces vLLM-compatible
INT4 output.
2026-06-01 04:22:07 +02:00
934be8ce48
fix: load tokenizer from base repo for quant venv compat
...
Merged model has tokenizer_class=TokenizersBackend (transformers 5.x)
which is unknown to transformers 4.57.6 in the quant venv.
2026-06-01 04:16:32 +02:00
0fa46c9fed
add AWQ quantization script (llm-compressor)
2026-06-01 04:15:15 +02:00
26e776db71
add substrate v5 training and build scripts
2026-06-01 03:52:55 +02:00
0e88c3c2ae
add v5 substrate training data (40 examples)
...
Manual hand-crafted examples for generic substrate LoRA:
- 01: identity/self-model
- 02-18: tool categories (159 tools grouped by function)
- 19-26: tricky tool disambiguation
- 27-32: framing (substrate vs persona, HMT, psychoacoustic trust)
- 33-40: EEMS deep-dive (architecture, recall, store, graph, legacy)
Pretty-printed JSON, compiled to JSONL via build pipeline.
System prompt uses generic assistant, not BT persona.
2026-05-31 14:32:19 +02:00
175d28a5aa
add bt7274 lora v2 and v3 adapter configs
2026-05-31 11:42:34 +02:00
cc7b5c88a3
add training scripts: v3, qwen3-8b, qlora-7b
2026-05-31 11:42:34 +02:00
0231a4d078
add bt7274 qwen3.5-27b lora v4 adapter config
2026-05-31 11:38:55 +02:00
00025dc804
add gonzales SD LoRA training images (LFS)
2026-05-31 11:38:49 +02:00
4cef9386b1
add docs: system lora plan, specialist specs, training review
2026-05-31 11:38:46 +02:00
4678816795
add training scripts: memory, specialist, mining, smoke test
2026-05-31 11:38:42 +02:00
df0d4a6eac
add training datasets: memory, persona, agent tools, serpent data
2026-05-31 11:38:37 +02:00
eee6aff617
enable git-lfs, track *.png
2026-05-31 11:38:25 +02:00
61d3e46ff1
add madcat-ml as submodule at ./docker
2026-05-31 11:31:02 +02:00
60f47d0379
fix: Qwen3.5 VL processor has no .encode() — use .tokenizer attr
2026-05-26 14:28:08 +02:00
5388df0075
fix: manual JSONL loader — pyarrow chokes on mixed tool_calls types
2026-05-26 14:20:37 +02:00
a562d753de
fix: adamw_8bit → adamw_torch — bitsandbytes has no cu132 binary
2026-05-26 12:10:30 +02:00
94515e7f6d
feat: bt7274 LoRA v4 — Hermes format, think blocks, 802 examples
2026-05-26 04:03:38 +02:00
122e73860b
feat: tts-norm LoRA — dataset generator + training script
...
gen_tts_dataset.py: 4960 synthetic examples, 22 categories (numbers,
currencies, dates, times, temperatures, acronyms, NATO phonetic, URLs,
markdown, etc). Bilingual EN/PL with explicit [lang] tag prefix.
train_tts_norm.py: Unsloth LoRA training for Qwen2.5-7B-Instruct.
Rank 16, 3 epochs, packing, max_seq 768. Trained on H100 in 20m38s,
final loss 0.091. Adapter: 154MB.
2026-05-26 00:14:51 +02:00
8137d278db
fix: parse tool_call arguments from JSON string to dict for Qwen3.5 template
2026-05-25 19:16:07 +02:00
4476520d5c
v3 dataset (582 ex, 14% direct) + Qwen3.5-27B training script
2026-05-25 18:57:17 +02:00
5deb9bd2b4
docs: bt7274 persona, specialist plan, tts-clean LoRA
2026-05-25 16:40:09 +02:00
9bcf1c2d9a
init: bt7274 lora training — v1 adapter, v2 dataset (500 examples), justfile
2026-05-25 16:37:30 +02:00
marauder-actual
madcat