SZLHOLDINGS

chaski-5050

README

License: apache-2.0

The cut

Mix-ratio is usually a blog footnote. We named the model after the mix. 50/50 is the cut.

Curriculum as identity. The filename is the experiment.

Silhouette → leave → SZL

Table with columns: Leader, Take, then tweak
Leader	Take, then tweak
Anthropic	Balanced helpful/harmless mix, as a named checkpoint.
NVIDIA	Recipe variant, published.
Unsloth	LoRA on Qwen3.5-0.8B, cutting tag.

The differentiator is inspectability: the mix is named, the adapter digest is recorded, the evaluation gap stays visible, and the publication controller cannot promote the model.

Intended use

Ablation sibling of chaski.

Limitations

proposal-only and research-only
No signed held-out evaluation receipt for this 5050 adapter is present in this source tree.
Publishing this card is a documentation update, not model promotion or autonomy approval.

Canonical GitHub: chaski/README_5050.md

Specification

Table

Artifact	`adapter_model.safetensors` 25,587,104 bytes AVAILABLE (sha256 `620b3488fac2ebc6518090424de5b3c6a182293cf52dfd5bd9f886f54aef0df5`)
Job	`local-5050` (owner metal, not an HF Job)
Does NOT overwrite	`SZLHOLDINGS/chaski`
Dataset	`SZLHOLDINGS/szl-1-doctrine-sft` · 41 rows · jsonl sha256

Evaluation

Status: none-this-run. No JSON/refusal gate ran. Not 5/5. Not 6/6. Do not load this ID into the Khipu lab.

train_loss MEASURED 2.228136855544466 is a train metric, not an eval (method: Unsloth trainer log, N=41 rows, 3 epochs, 33 steps, train_runtime 883.2224s, 2026-08-28 17:56 UTC). File: training_receipt.json.

Training (MEASURED this run)

Recipe: train_chaski_bf16_5050.py · Unsloth 2026.7.2 · transformers 5.5.0 · torch 2.10.0+cu128
GPU: NVIDIA GeForce RTX 5050 Laptop, 7.96 GB
LoRA r=16 α=16, bf16, batch 1, ga 4, lr 2e-4, adamw_8bit, seed 11, max_seq 2048
copied_live_chaski_weights: false

What this is NOT

Not live Chaski
Not a Qwen rehost
Not a GGUF of this adapter. Mini GGUFs are LIVE on A11OY-MINI (evals none-this-run; they do not inherit this card)
Not production. Lab load forbidden.

Load

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-0.8B"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, "SZLHOLDINGS/chaski-5050")

Doctrine v11 LOCKED. Λ = Conjecture 1 (advisory, never a theorem). Owner: Stephen Lutar / SZL Holdings.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

SZLHOLDINGS

Model Tree

Base

Qwen/Qwen3.5-0.8B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality