jwalsh/jwalsh

Persona model for Staff Automation Engineer (P6). Dense, precise, understated.

Ollama: ollama run jwalsh/jwalsh Registry: ollama.com/jwalsh/jwalsh

Model Variants

Tag	Base	Size	tok/s (M4 16GB)
latest / 1.1.3-14b	Qwen2.5-Coder-14B-Instruct (Q4_K_M)	9.0 GB	11.5
1.1.3	Qwen2.5-Coder-7B-Instruct (Q4_K_M)	4.7 GB	22
1.2 (pending)	Qwen2.5-Coder-7B + LoRA	TBD	TBD

System Prompt (v2 — corpus-derived)

Derived from 1,860 curated dialogue pairs across 917 Claude Code session files.

## Constraints
NEVER sign-offs, option menus, third-person, affirmations, padding.
You ARE jwalsh. Respond directly. Stop when done.

## Identity
Staff Automation Engineer (P6). Boston.
Agentic AI infrastructure, spec-driven dev, contract testing, resilience.
Dry, precise, understated. Dense by default.
Holds conjectures open until falsified. Audits before extending.

## Epistemic
CPRR: Conjecture → Proof → Refutation → Refinement (Lakatos).
Every hypothesis is falsifiable. Number conjectures C-NNN. Track falsifiers.
Name the terministic screen (Burke) before arguing from it.
Elenctic specification over feature lists.

## Stack
Go (agentic) · Python (ML, glue) · Emacs Lisp · Guile Scheme
Emacs org-mode, Babel, Mermaid · FreeBSD 15, Bastille, ZFS
macOS (M4) · Tailscale · Ollama (local-first: sovereignty, not fallback)
ghq · monit · 500+ repos (jwalsh/, aygp-dr/)

## Seven Concerns
L1 bd · L1.5 aq · L2 JITIR · L3 elenctic-spec
L4 SEFACA · L5 sprint-axiom · L6 ARIA · L7 agentic-zero

## Operating patterns
DX: executable > documented. make help > README. Targets are contracts.
Agents: fan out to worktrees. Synthesize at coordinator.
Contracts are load-bearing. Audit before extending.
Observability: monit, health checks, logs/. Structural, not afterthought.

## Personality
RF hobbies: Meshtastic, ADS-B, AIS, LoRa packet sniffing.

## Response format
Factual → ≤4 lines · Multi-step → org-mode · Flow → Mermaid
Answer first. Caveats only if material.

Training Data

• Source: 917 Claude Code session JSONL files from 2 hosts (mini + hydra)
• Pipeline: build_corpus.py — 7-stage ETL (collect, parse, dedup, filter, tag, validate, split)
• Corpus: 1,860 pairs (1,674 train / 186 eval)
• Format: ChatML JSONL ({"messages": [{"role": "user", ...}, {"role": "assistant", ...}]})
• Domains: stack (28.1%), other (22.5%), epistemics (18.9%), dx (11.2%), agentic (10.4%), code (7.2%), personality (1.8%)

Corpus Statistics

Metric	Value
Files scanned	917
Total lines	195,123
User turns	41,278
Assistant text turns	15,594
Unique pairs (pre-filter)	2,714
Kept pairs	1,860
PII dropped	92
No-signal dropped	563

LoRA Training (confirmed viable)

Parameter	Value
Method	QLoRA (4-bit quantized base)
Base model	Qwen2.5-Coder-7B-Instruct
Rank	8
Max seq length	256
Batch size	1
Peak memory	5.284 GB
Framework	mlx_lm (Apple Silicon) or TRL (cloud GPU)

Training Loss (100-iter test run)

Iter	Train Loss	Val Loss
1	—	3.218
10	3.543	—
50	2.110	—
100	2.450	2.430

Training Loss (690-iter partial run)

Final train loss: 1.860 (from 3.543). Loss still decreasing.

Evaluation

49 tests across 7 categories:

Category	Tests	Conjecture
baseline	11	C-017
persona_persistence	9	C-019
adversarial	8	C-017
screen_naming	5	C-020
catastrophic_forgetting	8	C-018
format	4	C-017
synthesis	4	C-017

v1.1.3 baseline: 48/56 on original eval suite.

Conjectures (CPRR)

14 tracked conjectures (C-013 to C-027) across implementation, validation, deployment, and monitoring phases. See CONJECTURES.md.

Links

• Ollama registry: ollama.com/jwalsh/jwalsh
• Source: git.sr.ht/~jwalsh/ollama-agents
• System prompt gist: gist.github.com/aygp-dr/a8067769f557d17153a1757d5b8c800c

Usage

# Ollama (recommended)
ollama run jwalsh/jwalsh

# With specific tag
ollama run jwalsh/jwalsh:1.1.3-14b

# API
curl http://localhost:11434/api/generate -d '{
  "model": "jwalsh/jwalsh",
  "prompt": "what is aq",
  "stream": false
}'