Kanha Qwen3 PIT checkpoint
Run identity
- Run ID:
kanha.ai-1.7b-pit-quality-v1 - Base model:
Qwen/Qwen3-1.7B - Base model revision:
70d244cc86ccca08cf5af4e1e306ecf908b1ad5e - Tokenizer revision:
70d244cc86ccca08cf5af4e1e306ecf908b1ad5e - Training method: PIT document continuation plus Q&A
- Final merged dtype:
bfloat16 - Source site: https://kanha.ai
- Training identity hash:
16baaea4d1c5a89dce8f3cdb27922bd9dbe51a1891b1b8bf5d90a3364406a7f7 - Train documents: 17
- Evaluation documents: 0
- Train Q&A pairs: 170
- Evaluation Q&A pairs: 0
Hyperparameters
- Max length:
2048 - Epochs:
20.0 - Learning rate:
1e-05 - Batch size:
8 - Grad accum:
2 - Seed:
42 - System prompt:
"You are a helpful assistant. Answer the user's question accurately and concisely." - Weight decay:
0.1 - Adam beta1:
0.9 - Adam beta2:
0.95 - Adam epsilon:
1e-08 - Warmup steps:
0 - Max grad norm:
1.0 - Bf16:
true - Tf32:
true - Gradient checkpointing:
true - Learning rate schedule:
"cosine_with_0.1_minimum" - Optimizer:
"adamw_torch" - Logging steps:
1 - Save strategy:
"epoch" - Save total limit:
3 - Data seed:
42 - Report to:
"none" - Remove unused columns:
false
Evaluation
- dates_recall:
1.0 - deterministic_pass_rate:
0.0 - list_recall:
0.07884615384615384 - numbers_recall:
0.7307692307692307 - refusal_rate:
0.0 - total:
26 - unsupported_value_rate:
0.6538461538461539 - urls_recall:
1.0
The external evaluation suite qualifies server-side behavior only. The exact converted artifact still requires browser and target-device validation.
MLC availability
No validated MLC artifact is included in this publication.
Limitations
The checkpoint can produce incorrect, incomplete, stale, or memorized content. The private training corpus is website-specific, and evaluation results do not establish general capability or production safety.
Provenance artifacts
research/run-manifest.jsonresearch/training-config.jsonresearch/publication-inventory.jsonresearch/evaluation/metrics.jsonresearch/evaluation/evaluation-manifest.json