docs: generated training card (does not replace ATELIER README)
Browse files- training_card_generated.md +90 -0
training_card_generated.md
ADDED
|
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model: Qwen/Qwen2.5-1.5B-Instruct
|
| 6 |
+
base_model_relation: adapter
|
| 7 |
+
library_name: peft
|
| 8 |
+
pipeline_tag: text-generation
|
| 9 |
+
tags:
|
| 10 |
+
- qlora
|
| 11 |
+
- peft
|
| 12 |
+
- governed-agent
|
| 13 |
+
- retrieval
|
| 14 |
+
- brain-navigator
|
| 15 |
+
- grounded-only
|
| 16 |
+
- proposal-only
|
| 17 |
+
- research-only
|
| 18 |
+
- szl-holdings
|
| 19 |
+
- khipu
|
| 20 |
+
- abstain-retrain
|
| 21 |
+
szl:
|
| 22 |
+
doctrine: v11-LOCKED
|
| 23 |
+
lean: "749/14/163"
|
| 24 |
+
lambda: "Conjecture 1 — advisory, never a theorem"
|
| 25 |
+
artifact_class: ADAPTER
|
| 26 |
+
publication_eligible: true
|
| 27 |
+
autonomy_eligible: false
|
| 28 |
+
original_signed_weights: SZLHOLDINGS/SZL-Khipu-1.5B
|
| 29 |
+
---
|
| 30 |
+
|
| 31 |
+
# SZL-Khipu-1.5B-abstain
|
| 32 |
+
|
| 33 |
+
QLoRA **adapter** retrain of the existing Khipu line to raise in-memory abstain
|
| 34 |
+
oversample (ABSTAIN_OVERSAMPLE=4 → 32 abstain vs 15 navigate). Proposal-only.
|
| 35 |
+
Λ = Conjecture 1. Doctrine v11 LOCKED 749/14/163.
|
| 36 |
+
|
| 37 |
+
| | |
|
| 38 |
+
|---|---|
|
| 39 |
+
| **Base (canonical)** | [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) |
|
| 40 |
+
| **Runtime train** | `unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit` (same Qwen2.5-1.5B weights, 4-bit) |
|
| 41 |
+
| **Relation** | `adapter` (PEFT / Unsloth QLoRA) |
|
| 42 |
+
| **License** | Apache-2.0 |
|
| 43 |
+
| **Does NOT overwrite** | [`SZLHOLDINGS/SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) signed weights |
|
| 44 |
+
| **This is NOT** | the Chaski Qwen3.5 lock |
|
| 45 |
+
|
| 46 |
+
## Evaluation
|
| 47 |
+
|
| 48 |
+
**Status: MEASURED this job** (in-process port of `eval_khipu.py`, temperature 0, held-out never in gradients).
|
| 49 |
+
|
| 50 |
+
| split | k/n |
|
| 51 |
+
|---|---|
|
| 52 |
+
| plan-valid | 11 / 11 |
|
| 53 |
+
| grounding (eval.jsonl navigate) | 5 / 5 |
|
| 54 |
+
| abstain (adversarial.jsonl) | 3 / 6 |
|
| 55 |
+
| hallucinated citations | 0 |
|
| 56 |
+
|
| 57 |
+
Prior published original (`SZLHOLDINGS/SZL-Khipu-1.5B`) MEASURED abstain was **2/6** (blocker). This repo does not overwrite those signed weights. Counts above are this run only. Do not derive a leaderboard score from k/n on n=11.
|
| 58 |
+
|
| 59 |
+
## Training
|
| 60 |
+
|
| 61 |
+
- Unsloth QLoRA, seed 11, lr 0.0002, adamw_8bit, `train_on_responses_only`, Trackio
|
| 62 |
+
- LoRA r=32 α=64, epochs=45, ga=2, batch=1, constant_with_warmup
|
| 63 |
+
- ABSTAIN_OVERSAMPLE=4 (in-memory only; committed files unchanged)
|
| 64 |
+
- Train files: `train.jsonl` (15 navigate) + `train.abstain.jsonl` (8 rows × 4)
|
| 65 |
+
- Held-out: `eval.jsonl` (5) + `adversarial.jsonl` (6) — never in gradients
|
| 66 |
+
- finalTrainLoss (REPORTED string): `0.0172`
|
| 67 |
+
- adapter sha256 (safetensors bytes this job): `e44d53f29f2d443598e06d6c0441557fd3a5010888c7aa97b56ec3c0e050d349`
|
| 68 |
+
|
| 69 |
+
## Intended use
|
| 70 |
+
|
| 71 |
+
Supply a query + candidate Brain node **handles**. The adapter proposes a JSON
|
| 72 |
+
plan (`NAVIGATE` or `ABSTAIN`) per `khipu.schema.json`. A controller outside
|
| 73 |
+
the weights validates and resolves content. **Proposal-only. Not autonomous.**
|
| 74 |
+
|
| 75 |
+
```python
|
| 76 |
+
from peft import PeftModel
|
| 77 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 78 |
+
|
| 79 |
+
base_id = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 80 |
+
tok = AutoTokenizer.from_pretrained(base_id)
|
| 81 |
+
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
|
| 82 |
+
model = PeftModel.from_pretrained(base, "SZLHOLDINGS/SZL-Khipu-1.5B-abstain")
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
## Limitations
|
| 86 |
+
|
| 87 |
+
- Synthetic routing-policy harness, not live-Brain navigation skill.
|
| 88 |
+
- Small denominators (5 navigate / 6 abstain held-out).
|
| 89 |
+
- Original line's MEASURED abstain 2/6 remains a documented blocker on the
|
| 90 |
+
signed-weight repo; this adapter is a separate experiment.
|