Downloads · 30 days
0
sadmanht/evalforge-mistral-7b-incident-diagnosis-qlora
evalforge-mistral-7b-incident-diagnosis-qlora is a text generation model from sadmanht. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
Model-card contract: phase10-huggingface-model-card-v1
Downloads · 30 days
0
Access
Public
Updated Sep 21, 2026
Repo size
168 MB
Likes
0
Public
Click a slice to open those files.
.safetensors168 MB · 97%
From the Hugging Face model README
Model-card contract: phase10-huggingface-model-card-v1
This repository publishes the PEFT adapter only for EvalForge's frozen FINETUNED arm. It is intended for reproducible research on the EvalForge production-incident diagnosis benchmark, not for autonomous production incident response.
sadmanht/evalforge-mistral-7b-incident-diagnosis-qloramistralai/Mistral-7B-Instruct-v0.3e8737b84b4470b28db3a0be719b362b1bd39a14de8ebf0c51d241516bd3c6bb44e476df6d412aaf6926705ecc53ca8cc3fcec065phase9-qlora-validation-v1b2f821a9cca2a54f3236d949a30ae480d5a7a511fe23335eca8be8ef7ee4873eb322d37e9f868b251fa14efb32c8412eab2ddaf3d5cd4adcf027c0f08d51411aevalforge-incident-diagnosis-v0.1.0zero-shot-baseline-v1root-cause-prediction-v1phase5-evaluator-v16064aaed457812a8102411eea47f02f78d812372097afe068ec53a540f9e1f7dThe adapter checkpoint was selected in Phase 09 using validation eval_loss before the Phase 10 locked test. The locked-test outcome was not used to retrain, switch checkpoints, change the prompt, change decoding, alter confidence scoring, or select a data-efficiency condition.
The benchmark contains six validation incidents and six locked-test incidents.
| Split / arm | Exact accuracy | Parse failure rate |
|---|---|---|
| Fine-tuned validation | 0.833333 | 0.166667 |
| Fine-tuned locked test | 0.833333 | 0.166667 |
| Zero-shot locked test | 1.000000 | 0.000000 |
On the six paired locked-test incidents, fine-tuned minus zero-shot exact accuracy was
-0.166667. The 95% paired bootstrap interval was
[-0.500000, 0.000000]
with 10000 resamples, and exact two-sided McNemar
p=1.000000.
The single fine-tuned locked-test miss was an INVALID_JSON parse failure after the
generation reached the frozen 128-token output limit. That observation is preserved without
post-test retuning.
Training uses only train incident families. Validation controls checkpoint/hyperparameter choices. Test families are locked and are never used as training examples. Synthetic descendants, when present, are permitted only for train families and retain parent-family lineage.
The Phase 10 data-efficiency matrix is secondary descriptive analysis only. It does not replace or select this primary adapter.
Load the exact base revision and then the adapter under the adapter/ directory with
Transformers + PEFT. EvalForge's scientific runtime uses bitsandbytes 4-bit NF4,
double quantization, float16 compute, deterministic decoding, and the frozen prompt/output
contract above.
The EvalForge repository preserves raw validation/test predictions, the exact adapter hash, paired statistics, frozen source/training identities, and the release evidence that records the immutable Hugging Face commit used for clean-download verification.