persona-model-trainer

Name: persona-model-trainer
Author: acnlabs

40stars5forks

How to install persona-model-trainer?

acnlabs/openpersona/persona-model-trainer

$npx -y skills add acnlabs/OpenPersona --skill persona-model-trainer

Installs into the current project.

Summary

Fine-tune any HuggingFace instruction-tuned model (Gemma 4, Qwen 3, Llama, Phi, Mistral, and more) on persona data from anyone-skill. Produces a self-contained, locally runnable persona model — no cloud API required.

Files · 1

View on GitHub

SKILL.md

1	# persona-model-trainer
2
3	Fine-tune a small local model on persona data (raw + distilled). Turn anyone-skill's output into a self-contained model that is the person — no prompting, no cloud, no latency.
4
5	Dependency chain: `anyone-skill` → `persona-knowledge` → `persona-model-trainer` → runnable persona model (`{model_id}`)
6
7	Input: `training/` folder produced by `anyone-skill` Step 6-D / `persona-knowledge` export (raw/ + conversations.jsonl + probes.json)
8	Output: LoRA/QLoRA adapter weights + GGUF / Ollama / vLLM / ONNX exports
9
10	> Full walkthrough: see `[references/pipeline-guide.md](references/pipeline-guide.md)` for the complete end-to-end guide (data → train → evaluate → version → run).
11
12	---
13
14	## When to use this skill
15
16	Trigger phrases:
17
18	- "train a model for this persona"
19	- "make it run locally / on my phone"
20	- "fine-tune on the distilled data"
21	- "I want a model, not just a prompt"
22	- "create a self-contained persona model"
23
24	Not suitable when:
25
26	- Effective assistant-role turns (raw/ + conversations.jsonl combined) < 200
27	- User only wants a quick prompt-based persona (use anyone-skill alone)
28
29	> Fictional characters and historical figures can be trained if `training/raw/` contains scripts, lore, speeches, or biographies — check actual turn count, not subject type.
30
31	---
32
33	## Quick Start — Pipeline Script
34
35	For standard use cases, `pipeline.sh` chains all phases (prepare → train → voice test → export) in one command:
36
37	```bash
38	# ── Gemma 4 preset (recommended for google/gemma-4-E4B-it) ──────────────────
39	# Apple Silicon — sets lora-rank=16, lora-layers=16, warmup-ratio=0.1, lora-alpha=16:
40	bash scripts/pipeline.sh \
41	--slug {slug} \
42	--model google/gemma-4-E4B-it \
43	--source ./training \
44	--method mlx \
45	--preset gemma4 \
46	--probes ./training/probes.json # optional: probe_score eval (generated by persona-knowledge)
47
48	# NVIDIA GPU — same preset, Unsloth backend (QLoRA, fits 8 GB VRAM):
49	bash scripts/pipeline.sh \
50	--slug {slug} \
51	--model unsloth/gemma-4-4b-it-bnb-4bit \
52	--source ./training \
53	--method unsloth \
54	--preset gemma4 \
55	--probes ./training/probes.json # omit if training/ was not exported by persona-knowledge
56
57	# ── Manual override (any model) ──────────────────────────────────────────────
58	# Local GPU — Apple Silicon (mlx) or NVIDIA (unsloth / qlora / lora):
59	bash scripts/pipeline.sh \
60	--slug {slug} \
61	--model {model_id} \
62	--source ./training \
63	--method mlx \
64	--lora-rank 16 \
65	--lora-layers 16 \
66	--warmup-ratio 0.05 \
67	--batch-size 2 \
68	--learning-rate 2e-4 \
69	--epochs 3
70
71	# No local GPU — train in Google Colab (free T4):
72	bash scripts/pipeline.sh \
73	--slug {slug} \
74	--model {model_id} \
75	--source ./training \
76	--method colab # generates colab_train_{slug}.ipynb, then exits
77	# → Upload .ipynb to colab.research.google.com → Run all → download adapter zip
78	# → Unzip into models/{slug}/export/ then:
79	bash scripts/pipeline.sh --slug {slug} --model {model_id} --source ./training \
80	--method skip-train # runs voice_test + export on the downloaded adapter
81
82	# Dry-run to validate setup (writes nothing):
83	bash scripts/pipeline.sh ... --dry-run
84
85	# After the script finishes, run the model with Ollama:
86	ollama create {slug} -f models/{slug}/export/ollama/Modelfile
87	ollama run {slug}
88
89	# Phase 8–9: bundle into installed persona pack
90	# --model-dir points to the version management root (BASE_DIR), not export/ directly
91	python scripts/pack_integrate.py \
92	--slug {slug} \
93	--model-dir models/{slug}/
94	# --pack-dir ~/.openpersona/personas/persona-{slug}/ # optional; auto-discovered if omitted
95	# → resolves export/ via manifest.json, copies artifacts, updates persona.json
96	```
97
98	Use the phases below for custom workflows, debugging, or when individual steps need tuning.
99
100	---
101
102	## Phase 1: Pre-flight Check
103
104	Read `training/metadata.json` (written by anyone-skill Step 6-D):
105
106	```json
107	{
108	"slug": "...",
109	"name": "...",
110	"subject_type": "personal \| public \| fictional \| historical \| archetype",
111	"source_count": 3,
112	"total_words": 48000,
113	"distilled_turns": 320,
114	"raw_files": ["whatsapp.jsonl", "essays.txt"],
115	"created_at": "2026-04-11T10:00:00Z"
116	}
117	```
118
119	**Gate — estimate effec

People who liked this also viewed

6 picks

Type

persona-model-trainer

40stars5forks

View on GitHub

How to install persona-model-trainer?

acnlabs/openpersona/persona-model-trainer

$npx -y skills add acnlabs/OpenPersona --skill persona-model-trainer

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# persona-model-trainer
2
3	Fine-tune a small local model on persona data (raw + distilled). Turn anyone-skill's output into a self-contained model that is the person — no prompting, no cloud, no latency.
4
5	Dependency chain: `anyone-skill` → `persona-knowledge` → `persona-model-trainer` → runnable persona model (`{model_id}`)
6
7	Input: `training/` folder produced by `anyone-skill` Step 6-D / `persona-knowledge` export (raw/ + conversations.jsonl + probes.json)
8	Output: LoRA/QLoRA adapter weights + GGUF / Ollama / vLLM / ONNX exports
9
10	> Full walkthrough: see `[references/pipeline-guide.md](references/pipeline-guide.md)` for the complete end-to-end guide (data → train → evaluate → version → run).
11
12	---
13
14	## When to use this skill
15
16	Trigger phrases:
17
18	- "train a model for this persona"
19	- "make it run locally / on my phone"
20	- "fine-tune on the distilled data"
21	- "I want a model, not just a prompt"
22	- "create a self-contained persona model"
23
24	Not suitable when:
25
26	- Effective assistant-role turns (raw/ + conversations.jsonl combined) < 200
27	- User only wants a quick prompt-based persona (use anyone-skill alone)
28
29	> Fictional characters and historical figures can be trained if `training/raw/` contains scripts, lore, speeches, or biographies — check actual turn count, not subject type.
30
31	---
32
33	## Quick Start — Pipeline Script
34
35	For standard use cases, `pipeline.sh` chains all phases (prepare → train → voice test → export) in one command:
36
37	```bash
38	# ── Gemma 4 preset (recommended for google/gemma-4-E4B-it) ──────────────────
39	# Apple Silicon — sets lora-rank=16, lora-layers=16, warmup-ratio=0.1, lora-alpha=16:
40	bash scripts/pipeline.sh \
41	--slug {slug} \
42	--model google/gemma-4-E4B-it \
43	--source ./training \
44	--method mlx \
45	--preset gemma4 \
46	--probes ./training/probes.json # optional: probe_score eval (generated by persona-knowledge)
47
48	# NVIDIA GPU — same preset, Unsloth backend (QLoRA, fits 8 GB VRAM):
49	bash scripts/pipeline.sh \
50	--slug {slug} \
51	--model unsloth/gemma-4-4b-it-bnb-4bit \
52	--source ./training \
53	--method unsloth \
54	--preset gemma4 \
55	--probes ./training/probes.json # omit if training/ was not exported by persona-knowledge
56
57	# ── Manual override (any model) ──────────────────────────────────────────────
58	# Local GPU — Apple Silicon (mlx) or NVIDIA (unsloth / qlora / lora):
59	bash scripts/pipeline.sh \
60	--slug {slug} \
61	--model {model_id} \
62	--source ./training \
63	--method mlx \
64	--lora-rank 16 \
65	--lora-layers 16 \
66	--warmup-ratio 0.05 \
67	--batch-size 2 \
68	--learning-rate 2e-4 \
69	--epochs 3
70
71	# No local GPU — train in Google Colab (free T4):
72	bash scripts/pipeline.sh \
73	--slug {slug} \
74	--model {model_id} \
75	--source ./training \
76	--method colab # generates colab_train_{slug}.ipynb, then exits
77	# → Upload .ipynb to colab.research.google.com → Run all → download adapter zip
78	# → Unzip into models/{slug}/export/ then:
79	bash scripts/pipeline.sh --slug {slug} --model {model_id} --source ./training \
80	--method skip-train # runs voice_test + export on the downloaded adapter
81
82	# Dry-run to validate setup (writes nothing):
83	bash scripts/pipeline.sh ... --dry-run
84
85	# After the script finishes, run the model with Ollama:
86	ollama create {slug} -f models/{slug}/export/ollama/Modelfile
87	ollama run {slug}
88
89	# Phase 8–9: bundle into installed persona pack
90	# --model-dir points to the version management root (BASE_DIR), not export/ directly
91	python scripts/pack_integrate.py \
92	--slug {slug} \
93	--model-dir models/{slug}/
94	# --pack-dir ~/.openpersona/personas/persona-{slug}/ # optional; auto-discovered if omitted
95	# → resolves export/ via manifest.json, copies artifacts, updates persona.json
96	```
97
98	Use the phases below for custom workflows, debugging, or when individual steps need tuning.
99
100	---
101
102	## Phase 1: Pre-flight Check
103
104	Read `training/metadata.json` (written by anyone-skill Step 6-D):
105
106	```json
107	{
108	"slug": "...",
109	"name": "...",
110	"subject_type": "personal \| public \| fictional \| historical \| archetype",
111	"source_count": 3,
112	"total_words": 48000,
113	"distilled_turns": 320,
114	"raw_files": ["whatsapp.jsonl", "essays.txt"],
115	"created_at": "2026-04-11T10:00:00Z"
116	}
117	```
118
119	**Gate — estimate effec

People who liked this also viewed

6 picks

Type