antonyamal071

crisisgraph-qwen3-4b-triage-lora-v3

README

License: apache-2.0

Artifact contents

adapter_model.safetensors — PEFT LoRA weights.
adapter_config.json — LoRA configuration and pinned base revision.
tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer artifacts saved with the training run.
prompt/triage-extraction-v2.1.1.txt — evaluated system prompt.
schema/triage-extraction-v2.json — strict structured-output contract.
gguf/crisisgraph-qwen3-4b-triage-lora-v3-f16.gguf — F16 LoRA adapter for llama.cpp; it still requires a compatible Qwen3-4B-Instruct-2507 base GGUF.
release-manifest.json and SHA256SUMS — provenance and integrity data.

Base model

Table with columns: Field, Value
Field	Value
Base model	`Qwen/Qwen3-4B-Instruct-2507`
Base revision	`cdbee75f17c01a7cc42f958dc650907174af0554`
Relationship	PEFT LoRA adapter

The base model is not redistributed in this repository.

Intended use

The adapter is intended for controlled experimentation with extraction of:

victim or pickup location;
headcount;
explicitly requested response asset;
named road hazards and reported duration;
uncertainty and the need for human review.

The evaluated output is governed by the included prompt and JSON schema. Using a different prompt, schema, decoding configuration, base revision, or adapter scale is a different system configuration and is not covered by the reported results.

Training

The adapter was trained with QLoRA on a Tesla T4 for two epochs.

Table with columns: Setting, Value
Setting	Value
Training examples	300
Validation examples	72
Data origin	AI-assisted synthetic scenarios
Maximum sequence length	1,024
LoRA rank / alpha	16 / 32
LoRA dropout	0.05
Target modules	All linear projection modules
Learning rate	`1e-4`
Micro-batch / gradient accumulation	1 / 8
Training runtime	1,566.28 seconds

The synthetic data covers clear, missing, contradictory, hazard, adversarial, irrelevant, Malayalam, and Manglish cases. Records received an AI-assisted first-pass semantic review. An independent second review was not completed, so the data and adapter must not be described as independently reviewed.

Teacher-forced validation loss was 0.01027 and teacher-forced token accuracy was 0.99685. Those training metrics are not extraction accuracy and do not establish real-world generalization.

Frozen regression evaluation

The adapter was evaluated on the existing 120-case CrisisGraph regression suite using prompt triage-extraction-v2.1.1, strict JSON-schema decoding, temperature zero, and a Q4_K_M GGUF of the pinned base model through llama.cpp.

Table with columns: Metric, V3 result
Metric	V3 result
Schema validity	100.0%
Critical-field accuracy	94.0%
Exact match across five critical fields	81.7% (98/120)
Raw human-review recall	88.2% (60/68)
Raw human-review precision	95.2%
Cases with evaluator-detected unsupported facts	14.2% (17/120)
Inference errors	0

This suite was used repeatedly during development and is not a blind holdout. The adapter failed the predeclared clear-case exact-match gate with 18/20 rather than the required 20/20. It produced eight review false negatives and three review false positives. V3 is therefore the leading experimental CrisisGraph adapter, not an unconditionally promoted model.

Loading with Transformers and PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-4B-Instruct-2507"base_revision = "cdbee75f17c01a7cc42f958dc650907174af0554"adapter_id = "antonyamal071/crisisgraph-qwen3-4b-triage-lora-v3" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base = AutoModelForCausalLM.from_pretrained(    base_id,    revision=base_revision,    torch_dtype="auto",    device_map="auto",)model = PeftModel.from_pretrained(base, adapter_id)model.eval()

Use the included system prompt and JSON schema when reproducing the structured extraction configuration.

Loading with llama.cpp

The GGUF file in gguf/ is an adapter, not a complete GGUF model:

bash

llama-server \  --model /path/to/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \  --lora /path/to/crisisgraph-qwen3-4b-triage-lora-v3-f16.gguf \  --port 8000

Use a compatible GGUF conversion of the same base model. Reproduction details for the measured CPU run are available in the CrisisGraph repository.

Limitations

Training and validation data are synthetic and narrowly task-specific.
The regression suite is exposed, not a blind generalization test.
Real Kerala emergency-call or field-report distributions were not available.
Malayalam and Manglish coverage is limited and cannot represent all dialects, spelling variation, code-switching, or operational language.
The adapter may invent missing facts, choose one side of a contradiction, or over-escalate a complete report.
Model confidence values are task-policy labels, not calibrated probabilities.
Structured-output validity does not imply semantic correctness.

Any downstream system should validate the schema, apply deterministic missing and contradiction guards, preserve the raw report, record model provenance, and require human authority for uncertain facts and operational hazards.

License

This adapter is released under Apache 2.0, consistent with the base model's published license. Users must also review and comply with the base model's terms and all laws and policies applicable to their deployment.

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

antonyamal071

Model Tree

Base

Qwen/Qwen3-4B-Instruct-2507

Adapter

this model

Input Modalities

Text

Output Modalities