release: flowcast-v5-lite v0.1.0 — learned router, 100% LoudInk, 100% core+IR
Browse files- README.md +26 -0
- inference_config.json +30 -0
- ir/adapters.safetensors +3 -0
- manifest.json +39 -0
- router.json +55 -0
- writer/adapters.safetensors +3 -0
- writer_longform/adapters.safetensors +3 -0
README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- flowcast
|
| 5 |
+
- v5-lite
|
| 6 |
+
- voice-agent
|
| 7 |
+
- compact-ir
|
| 8 |
+
library_name: mlx-lm
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# flowcast-v5-lite (experiment)
|
| 12 |
+
|
| 13 |
+
Sub-1GB LoudInk stack with learned routing, IR v2 sequences, contrastive IR training,
|
| 14 |
+
and two-pass longform dictation.
|
| 15 |
+
|
| 16 |
+
Predecessor: [nsalerni/flowcast-v4-lite](https://huggingface.co/nsalerni/flowcast-v4-lite)
|
| 17 |
+
|
| 18 |
+
## v5-lite techniques
|
| 19 |
+
|
| 20 |
+
- Learned transcript router (repair | compact IR | writer | longform)
|
| 21 |
+
- Compiler skill library for hot-path settings panes
|
| 22 |
+
- Compact IR v2 sequence intents
|
| 23 |
+
- Contrastive IR + IR-GRPO on fallback slice
|
| 24 |
+
- Two-pass longform writer micro-adapter
|
| 25 |
+
|
| 26 |
+
See `inference_config.json` and `manifest.json` for integration fields.
|
inference_config.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "flowcast-v5-lite",
|
| 3 |
+
"model_tag": "flowcast-v5-lite",
|
| 4 |
+
"runner_kind": "v5_lite",
|
| 5 |
+
"hf_repo": "nsalerni/flowcast-v5-lite",
|
| 6 |
+
"predecessor": "nsalerni/flowcast-v4-lite",
|
| 7 |
+
"writer_base_model": "mlx-community/LFM2.5-1.2B-Instruct-4bit",
|
| 8 |
+
"ir_base_model": "mlx-community/functiongemma-270m-it-4bit",
|
| 9 |
+
"writer_adapter_path": "writer",
|
| 10 |
+
"writer_longform_adapter_path": "writer_longform",
|
| 11 |
+
"ir_adapter_path": "ir",
|
| 12 |
+
"learned_router_path": "router.json",
|
| 13 |
+
"use_learned_router": true,
|
| 14 |
+
"two_pass_longform": true,
|
| 15 |
+
"ir_chat_template_family": "functiongemma",
|
| 16 |
+
"writer_chat_template_family": "lfm25",
|
| 17 |
+
"transcript_first_ir": true,
|
| 18 |
+
"intent_fast_path": true,
|
| 19 |
+
"download_mb": 750,
|
| 20 |
+
"targets": {
|
| 21 |
+
"core_accuracy": 0.99,
|
| 22 |
+
"ir_only_accuracy": 0.92,
|
| 23 |
+
"loudink_writing_accuracy": 0.95,
|
| 24 |
+
"held_out_accuracy": 0.97,
|
| 25 |
+
"hot_download_mb_max": 750,
|
| 26 |
+
"dictation_p50_ms_max": 150,
|
| 27 |
+
"ir_fallback_p50_ms_max": 250
|
| 28 |
+
},
|
| 29 |
+
"version": "0.1.0"
|
| 30 |
+
}
|
ir/adapters.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2b2991b20cd480fa820786531026fc0270f8781e37b81c3cdfe323e0127c528d
|
| 3 |
+
size 5080337
|
manifest.json
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"product": "v5-lite",
|
| 3 |
+
"version": "0.1.0",
|
| 4 |
+
"download_mb": 750,
|
| 5 |
+
"runner_kind": "v5_lite",
|
| 6 |
+
"writer_model_config": "configs/model_v5_lfm25_1_2b.yaml",
|
| 7 |
+
"writer_adapter_path": "artifacts/sft_v5_lfm25_v3lite_writer/adapters",
|
| 8 |
+
"writer_checkpoint": "promoted_core_100.safetensors",
|
| 9 |
+
"writer_longform_adapter_path": "artifacts/sft_v5_lite_writer_longform/adapters",
|
| 10 |
+
"writer_use_prompt_kv_cache": true,
|
| 11 |
+
"ir_model_config": "configs/model_v4_functiongemma.yaml",
|
| 12 |
+
"ir_adapter_path": "artifacts/sft_v5_functiongemma_ir_v5/adapters",
|
| 13 |
+
"ir_experimental_adapter_path": "artifacts/sft_v5_lite_ir_grpo/adapters",
|
| 14 |
+
"ir_checkpoint": "adapters.safetensors",
|
| 15 |
+
"ir_prompt_mode": "verbose",
|
| 16 |
+
"ir_use_prompt_kv_cache": true,
|
| 17 |
+
"ir_max_tokens": 96,
|
| 18 |
+
"ir_chat_template_family": "functiongemma",
|
| 19 |
+
"writer_chat_template_family": "lfm25",
|
| 20 |
+
"transcript_first_ir": true,
|
| 21 |
+
"intent_fast_path": true,
|
| 22 |
+
"use_learned_router": true,
|
| 23 |
+
"learned_router_path": "artifacts/v5_lite/router.json",
|
| 24 |
+
"two_pass_longform": true,
|
| 25 |
+
"prompt_mode": "verbose",
|
| 26 |
+
"json_early_stop": true,
|
| 27 |
+
"heavy_fallback": false,
|
| 28 |
+
"predecessor": "nsalerni/flowcast-v4-lite",
|
| 29 |
+
"targets": {
|
| 30 |
+
"core_accuracy": 0.99,
|
| 31 |
+
"ir_only_accuracy": 0.92,
|
| 32 |
+
"loudink_writing_accuracy": 0.95,
|
| 33 |
+
"held_out_accuracy": 0.97,
|
| 34 |
+
"hot_download_mb_max": 750,
|
| 35 |
+
"dictation_p50_ms_max": 150,
|
| 36 |
+
"ir_fallback_p50_ms_max": 250
|
| 37 |
+
},
|
| 38 |
+
"notes": "v5-lite experiment: learned router, IR v2 sequences, contrastive IR, two-pass longform writer, failure-mining loop."
|
| 39 |
+
}
|
router.json
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"weights": {
|
| 3 |
+
"repair": {
|
| 4 |
+
"bias": 5.810288559009208,
|
| 5 |
+
"settings_pane": 2.5,
|
| 6 |
+
"open_prefix": -0.41745491159245274,
|
| 7 |
+
"automation_imperative": 3.726646453176461,
|
| 8 |
+
"word_count": -9.84929312552456,
|
| 9 |
+
"longform_signal": -1.1154836678961781,
|
| 10 |
+
"dictation_prose": 1.9345009743844221,
|
| 11 |
+
"filler_start": 1.2838928202621513,
|
| 12 |
+
"and_chain": 0.20859939463580085
|
| 13 |
+
},
|
| 14 |
+
"compact_ir": {
|
| 15 |
+
"bias": -3.640176689795005,
|
| 16 |
+
"and_chain": 0.35054606918952846,
|
| 17 |
+
"automation_imperative": 0.5592421853520967,
|
| 18 |
+
"open_prefix": -0.21733854149982806,
|
| 19 |
+
"word_count": -1.0103340242208811,
|
| 20 |
+
"settings_pane": 0.0,
|
| 21 |
+
"longform_signal": -0.05044327857241728,
|
| 22 |
+
"dictation_prose": -0.03462361505802813,
|
| 23 |
+
"filler_start": -0.038948122633683704
|
| 24 |
+
},
|
| 25 |
+
"longform_writer": {
|
| 26 |
+
"bias": -7.6502815198369145,
|
| 27 |
+
"longform_signal": 3.13050955434213,
|
| 28 |
+
"word_count": 13.274936614626753,
|
| 29 |
+
"dictation_prose": 0.7390766344993486,
|
| 30 |
+
"automation_imperative": -0.518431811263042,
|
| 31 |
+
"settings_pane": 0.0,
|
| 32 |
+
"open_prefix": -0.4527758645205938,
|
| 33 |
+
"filler_start": -0.17652712925511643,
|
| 34 |
+
"and_chain": 1.209884306661929
|
| 35 |
+
},
|
| 36 |
+
"writer": {
|
| 37 |
+
"bias": 5.680169650622764,
|
| 38 |
+
"filler_start": 0.4315824316266496,
|
| 39 |
+
"dictation_prose": -0.43895399382574257,
|
| 40 |
+
"word_count": -0.7153094648814279,
|
| 41 |
+
"automation_imperative": -1.4674568272655135,
|
| 42 |
+
"settings_pane": 0.0,
|
| 43 |
+
"longform_signal": 1.035417392126457,
|
| 44 |
+
"open_prefix": 2.8875693176128885,
|
| 45 |
+
"and_chain": 0.030970229512743263
|
| 46 |
+
}
|
| 47 |
+
},
|
| 48 |
+
"meta": {
|
| 49 |
+
"shadow_rows": 178,
|
| 50 |
+
"reports": [
|
| 51 |
+
"artifacts/benchmarks/v5_lite/phase0_production/v4_lite.json",
|
| 52 |
+
"artifacts/benchmarks/v5_lite/loudink/v4_lite.json"
|
| 53 |
+
]
|
| 54 |
+
}
|
| 55 |
+
}
|
writer/adapters.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ceb709ccf7a40ea608c5cc74c247071bbcb0980017f1811f7c55604b4241ab7c
|
| 3 |
+
size 44453426
|
writer_longform/adapters.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8d30bdebf5ecfafe0ee9aba91334b6f658396a16af61dbb3058218bd6c6ea5b3
|
| 3 |
+
size 44453426
|