japhba commited on
Commit
03d2fbd
·
verified ·
1 Parent(s): 269a342

Update model card for failure_mode_enum

Browse files
Files changed (1) hide show
  1. README.md +65 -0
README.md ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: Qwen/Qwen3-14B
3
+ library_name: peft
4
+ pipeline_tag: text-generation
5
+ tags:
6
+ - lora
7
+ - peft
8
+ - qwen3
9
+ - multitier-lora
10
+ - tier-2
11
+ - failure-mode-enum
12
+ ---
13
+
14
+ # multitier-lora-tier2-failure-mode-enum
15
+
16
+ This repository contains the best-selected LoRA adapter for `failure_mode_enum` from the multitier LoRA sweep.
17
+
18
+ ## Model Details
19
+
20
+ - Base model: `Qwen/Qwen3-14B`
21
+ - Tier: `2` (Procedure following)
22
+ - LoRA identifier: `failure_mode_enum`
23
+ - Teacher model used for synthetic data generation: `gemini-3.1-flash-lite-preview`
24
+ - Training objective: Supervised fine-tuning on synthetic procedure-execution demonstrations.
25
+ - Best checkpoint metric: `procedure_margin = 0.6070` at step `141`
26
+ - Training rows used: `1500`
27
+ - Evaluation rows used: `100`
28
+
29
+ ## Intended Behavior
30
+
31
+ Systematically enumerate the ways a system or plan can fail, then assess which failure modes matter most by severity, likelihood, and detectability.
32
+
33
+ - Meta domains: tech_systems, product_design, business_ops, public_policy, research_agenda, org_strategy
34
+ - Output structure: ## System, ## Failure Modes, ## Severity, ## Likelihood, ## Detectability, ## Controls
35
+
36
+ ## Files
37
+
38
+ - `adapter_model.safetensors` and `adapter_config.json`: the selected PEFT adapter weights
39
+ - `best_checkpoint.json`: the selected evaluation checkpoint metadata
40
+ - `eval_log.jsonl`: evaluation history across checkpoints
41
+ - `run_config.jsonl`: the training configuration used for this run
42
+
43
+ ## Dataset Link
44
+
45
+ The companion synergy Q&A dataset is published at:
46
+
47
+ - https://huggingface.co/datasets/mats-10-sprint-cs-jb/multitier-lora-synergy-qa
48
+
49
+ ## Usage
50
+
51
+ ```python
52
+ from peft import PeftModel
53
+ from transformers import AutoModelForCausalLM, AutoTokenizer
54
+
55
+ base_id = "Qwen/Qwen3-14B"
56
+ adapter_id = "mats-10-sprint-cs-jb/multitier-lora-tier2-failure-mode-enum"
57
+
58
+ tokenizer = AutoTokenizer.from_pretrained(adapter_id)
59
+ model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
60
+ model = PeftModel.from_pretrained(model, adapter_id)
61
+ ```
62
+
63
+ ## Notes
64
+
65
+ These adapters were selected from the `best_adapter` checkpoint for each LoRA rather than from the final training step. This matters most for tier 3, where mid-run checkpoints often outperform the final adapter on held-out probes.