Modelpublic

HarleyCooper/Laguna-XS.2-Adaption-Dakota-QA-GRPO

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes

Model Card

Laguna-XS.2-Adaption-Dakota-QA-GRPO

This repository documents a completed Prime Intellect Hosted Training RL run on the Adaption Labs Dakota-English QA dataset.

Important: at initial publication time this repository is a run card and artifact ledger, not a downloadable PEFT adapter package. Prime lists READY adapters for the run, but the Prime CLI exposes them as deployment targets and R2 checkpoint locations rather than direct Hugging Face weight files. The final adapter deployment attempt returned DEPLOY_FAILED; the step-75 adapter was still DEPLOYING after repeated polling. Weight files will be added here if/when an export path is available.

Current status

—HF repo: HarleyCooper/Laguna-XS.2-Adaption-Dakota-QA-GRPO
—Training platform: Prime Intellect Hosted Training
—Prime run: `bbu5xvdv42zh8o6vp955klhy`
—Smoke run: `l8bs3nhaqitribgit4xoyunu`
—Base model: poolside/Laguna-XS.2
—Dataset: `HarleyCooper/adaption-dakota-english-qa`
—Prime verifier environment: harleycooper/adaption-dakota-qa
—Run status: COMPLETED
—Started: 2026-06-04 03:12 UTC
—Completed: 2026-06-04 06:41 UTC
—Cost: $0.00 under Prime's free Laguna offer
—Tokens processed: 3.96M
—Samples processed: 12,800
—Problems processed: 1,595

Training configuration

Setting	Value
max steps	100
completed latest step	99
batch size	128
rollouts per example	8
max completion tokens	768
learning rate	`1e-4`
env dataset limit	full dataset, `max_examples = -1`
eval fraction	0.10
eval examples	256

The launch config is included in this repo as `training_config/laguna-full-free.toml`.

Result summary

Metric	Step 0	Step 99
reward mean	0.283300	0.432624
reward max	0.502806	0.652041
character-F1 reward	0.326565	0.635267
Dakota-term reward	0.255516	0.420664
keyword reward	0.452124	0.367710
length reward	0.571849	0.924520
zero-advantage filter	0.0	0.0
empty rollouts	0.0	0.0
errored rollouts	0.0	0.0

The main result is that the custom Dakota QA reward produced non-degenerate signal over the full run: reward mean rose from 0.283 to 0.433, character F1 rose from 0.327 to 0.635, Dakota-term reward rose from 0.256 to 0.421, and Prime reported no empty rollouts, no errored rollouts, and no zero-advantage filtering at the final step.

Reward function

The Prime Verifiers environment scores single-turn Dakota-English QA completions with a deterministic continuous reward:

Component	Weight	Purpose
exact normalized match	20%	Rewards exact target-answer reproduction when it happens
Dakota term coverage	30%	Rewards correct Dakota forms and orthography
character F1	25%	Rewards close overlap with the reference answer
keyword coverage	15%	Rewards semantic/lexical overlap with target answer keywords
length/readability	10%	Penalizes empty, extremely short, or rambling responses

This is deliberately not all-or-nothing. RL needs variance across rollouts, and natural-language dictionary/grammar answers can be partially correct.

Checkpoints and adapters

Prime reported these READY checkpoints/adapters:

Step	Checkpoint ID	Adapter ID	Status	Size
50	`h0y0un58kea8uro0kqkv1zuz`	`djzuko6ckwuyfgj6p9c6g10s`	READY	18.5 GB
75	`xq30oegs88b6jgk2l77sx1qb`	`h1rwu671te8cmng5rw2p24vf`	READY	18.5 GB
final	—	`w1n9h644zkiiqjef1ux6jjr4`	READY in deployments list	—

Prime checkpoint storage paths are recorded in `analysis/run_summary.json`. They are Prime R2 URLs, not public download URLs.

Inference status

Prime inference is live for the step-75 adapter:

text

poolside/Laguna-XS.2:h1rwu671te8cmng5rw2p24vf

Working Prime CLI query from WSL:

bash

prime inference chat \
  "poolside/Laguna-XS.2:h1rwu671te8cmng5rw2p24vf" \
  "Given the verb a-kpa'-spa, what are the forms for I suffer patiently and You suffer patiently?" \
  --max-tokens 25000

Observed status:

—Step-75 adapter h1rwu671te8cmng5rw2p24vf: inference works through Prime Inference.
—Final adapter w1n9h644zkiiqjef1ux6jjr4: earlier deployment attempt returned DEPLOY_FAILED; querying it returned model-not-found.
—Direct local weight-file inference is still blocked until Prime exposes/export-downloads the LoRA adapter weights. Current inference is remote Prime-hosted inference from a local WSL terminal.

Initial manual probe note: the adapter responds, but the first checked a-kpa'-spa answer was not yet linguistically correct against the dataset target. Treat this as an available research checkpoint that needs eval, not an authoritative Dakota assistant.

Included files

—`analysis/run_summary.json`: machine-readable run/card summary
—`analysis/prime_metrics_compact.csv`: compact per-step metric table
—`examples/rollouts_step90.md`: human-readable sample completions
—`examples/rollouts_step90.json`: raw Prime rollout sample page
—`training_config/laguna-full-free.toml`: launch config

Intended use

This run is intended for research on verifier-driven RL for historical Dakota-English dictionary and grammar QA. It is useful for studying whether dense, source-derived reward functions can improve Dakota orthography preservation, Dakota term inclusion, and concise dictionary-style answering.

It is not an authoritative Dakota language tool, a replacement for community language expertise, or a production translation system.

Ethical and source notes

The dataset and reward are derived from historical Dakota-English dictionary material and generated QA structure. Historical sources can contain colonial-era framing, extraction errors, orthographic inconsistency, and gaps relative to contemporary language use. Any Dakota-language output should be reviewed with appropriate linguistic and community expertise.