Upload README.md

README.md

@@ -1,101 +1,134 @@
 ---
+base_model: togethercomputer/gpt-oss-20b-bf16
+library_name: peft
+license: cc-by-4.0
 language:
 - ha
 - yo
 - sw
-license: cc-by-4.0
-task_categories:
-- text-classification
 tags:
 - sentiment-analysis
 - african-languages
-- hausa
-- yoruba
-- swahili
+- lora
+- peft
 - autoscientist-challenge
-size_categories:
-- 10K<n<100K
 ---
 
-# Combined Sentiment Dataset: Hausa, Yorùbá, Swahili
-
-A stitched multi-source sentiment classification dataset combining three
-independently collected sentiment corpora for Hausa, Yorùbá, and Swahili,
-built for the [Adaption Labs AutoScientist Challenge](https://adaptionlabs.ai/blog/autoscientist-challenge)
-(Language category).
-
-## Why this dataset
-
-Existing sentiment resources for these languages are dominated by a single
-domain (Twitter). This dataset combines **three different sources across
-three different domains** to reduce domain overfitting and give a more
-robust sentiment signal:
-
-| Source | Domain | Languages | Rows |
-|---|---|---|---|
-| [AfriSenti](https://github.com/afrisenti-semeval/afrisent-semeval-2023) | Twitter | Hausa, Yorùbá, Swahili | 40,290 |
-| [NollySenti](https://github.com/IyanuSh/NollySenti) | Nollywood movie reviews (human-translated) | Hausa, Yorùbá | 2,510 |
-| [Neurotech-HQ Swahili](https://github.com/Neurotech-HQ/swahili-sentiment-analysis-dataset) | Social media / product reviews (back-translated) | Swahili | 3,925 |
-
-## Dataset structure
-
-- `text`: the raw text (tweet, movie review, or social media comment)
-- `label`: one of `positive`, `negative`, `neutral` (single unified label column)
-- `language`: `hausa`, `yoruba`, or `swahili`
-- `source`: which of the three original datasets the row came from
-- `domain`: `twitter`, `movie_review`, or `social_media_reviews`
-- `split`: `train`, `dev`, or `test`
-
-## Splits
-
-All three languages use an identical **70 / 15 / 15** train/dev/test split,
-stratified by label, computed after pooling all sources per language
-(NOT the original per-source splits — see Limitations).
-
-| Language | Total | Train | Dev | Test |
-|---|---|---|---|---|
-| Hausa | 23,162 | 16,213 | 3,474 | 3,475 |
-| Yorùbá | 16,627 | 11,639 | 2,494 | 2,494 |
-| Swahili | 6,936 | 4,855 | 1,040 | 1,041 |
+# African Languages Sentiment Classifier (Hausa, Yorùbá, Swahili)
+
+> ## ⚠️ Read before using this model
+> This is a **LoRA adapter produced by an automated fine-tuning/data-evolution
+> pipeline (Adaption Labs' AutoScientist)**, released as a submission for the
+> [AutoScientist Challenge](https://adaptionlabs.ai/blog/autoscientist-challenge).
+> It is a **research artifact, not a vetted production model**. Before you use
+> it, please understand what it actually is:
+>
+> - It was fine-tuned on a **machine-evolved version of the training data**
+>   (AutoScientist's "Adaptive Data" step rewrote/expanded the original
+>   labeled examples into `enhanced_prompt` / `enhanced_completion` pairs
+>   with synthetic reasoning traces). It was **not** trained directly on the
+>   raw human-annotated dataset as-is.
+> - The **evaluation numbers below come from AutoScientist's internal
+>   judge-based quality and win-rate metrics**, not standard classification
+>   accuracy, precision, recall, or F1. Treat "Grade" and "Win Rate" as
+>   AutoScientist's own scoring rubric, not an external, independently
+>   reproducible benchmark.
+> - **General-capability win rate slightly regressed** after adaptation
+>   (52 → 48, see below) — this model appears to trade a small amount of
+>   general capability for task-specific sentiment performance. That's a
+>   normal fine-tuning tradeoff, not a bug, but you should know it's there.
+> - This adapter has **not been independently evaluated against the
+>   `Davlan/afrisenti-twitter-sentiment-afroxlmr-large` baseline** on a
+>   standard accuracy/macro-F1 metric. If you need that comparison, run it
+>   yourself before drawing conclusions about relative performance.
+>
+> In short: understand the pipeline behind this artifact before you plug it
+> into anything downstream.
+
+## Model Details
+
+- **Base model:** `togethercomputer/gpt-oss-20b-bf16`
+- **Adapter type:** LoRA (PEFT), rank 64, alpha 128, target modules `q_proj`/`k_proj`/`v_proj`/`o_proj`
+- **Task formulation:** causal LM, prompt → single-token/word completion (average prompt length 15.56 words, completion length 1 word) — i.e. the model is trained to generate the sentiment label as its next-token completion, not as a traditional classification head
+- **Languages:** Hausa, Yorùbá, Swahili
+- **License:** CC-BY-4.0
+- **Produced via:** [Adaption Labs AutoScientist](https://adaptionlabs.ai/blog/autoscientist-challenge) (Language category submission)
+
+## Training Data
+
+Trained on an AutoScientist-evolved version of the companion dataset
+[`gospelgit/African-Languages_Sentiments`](https://huggingface.co/datasets/gospelgit/African-Languages_Sentiments),
+which itself combines AfriSenti (Twitter), NollySenti (Nollywood movie
+reviews), and the Neurotech-HQ Swahili sentiment dataset (social
+media/product reviews).
+
+- **Adapted dataset size:** 15,280 rows (AutoScientist's "Adaptive Data" step
+  rewrote and filtered the original ~46,725-row combined corpus into
+  enhanced prompt/completion pairs — see disclaimer above)
+- Labels: `positive`, `negative`, `neutral`
+
+## Training Procedure
+
+- 5 epochs, 585 total steps
+- Train/eval loss decreased steadily across all 5 epochs (eval loss:
+  0.828 → 0.787 → 0.769 → 0.760 → 0.757)
+- Learning rate: warm-up then decay schedule
+- Framework: PEFT 0.15.1
+
+## Evaluation (AutoScientist internal metrics)
+
+These are AutoScientist's own judge-based scores comparing the base model
+against the fine-tuned ("adapted") model — **not standard accuracy/F1**:
+
+| Metric | Before (base) | After (adapted) |
+|---|---|---|
+| Quality score (0–10 scale) | 3.0 | 6.9 (+130% relative) |
+| Grade | E | C |
+| Percentile | 1.3 | 8.4 |
+| Win rate — on this dataset | 44 | 57 |
+| Win rate — general category (all tasks) | 52 | 48 |
+
+**Read this table carefully**: task-specific quality improved substantially
+(grade E→C, +130% relative quality score), but the general-category win rate
+slightly *dropped* (52→48), meaning the adaptation traded a small amount of
+general-purpose capability for sentiment-task performance. This is disclosed
+deliberately — don't assume "adapted" is strictly better in every dimension.
+
+## Intended Use
+
+Sentiment classification (positive/negative/neutral) for short-form text in
+Hausa, Yorùbá, or Swahili, primarily for research and benchmarking purposes
+within the AutoScientist Challenge. Not validated for production deployment.
+
+## How to Use
+
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base_model = AutoModelForCausalLM.from_pretrained("togethercomputer/gpt-oss-20b-bf16")
+model = PeftModel.from_pretrained(base_model, "gospelgit/African-Languages-Sentiment-Classifier")
+tokenizer = AutoTokenizer.from_pretrained("gospelgit/African-Languages-Sentiment-Classifier")
+
+prompt = "Classify the sentiment of this text as positive, negative, or neutral: <your text here>"
+inputs = tokenizer(prompt, return_tensors="pt")
+output = model.generate(**inputs, max_new_tokens=5)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
 
 ## Limitations
 
-- **Label imbalance across sources**: NollySenti and the Neurotech Swahili
-  set are binary (positive/negative only). `neutral` examples come
-  exclusively from AfriSenti, so neutral coverage is thinner relative to
-  positive/negative for all three languages.
-- **Re-split, not original splits**: because splits were recomputed by
-  pooling sources, this is not directly comparable row-for-row to
-  benchmarks trained on the original AfriSenti or NollySenti splits (e.g.
-  `Davlan/afrisenti-twitter-sentiment-afroxlmr-large`). Use as a fresh
-  baseline, not a drop-in replacement.
-- **Neurotech Swahili has no official original split** — an 80/10/10 split
-  was assigned with a fixed seed (42) before being pooled and re-split here.
-- **Swahili domain composition differs** from Hausa/Yorùbá: it is roughly
-  half Twitter, half social-media/product reviews, while Hausa and Yorùbá
-  are Twitter-dominant with a smaller movie-review slice.
+- Evaluated via AutoScientist's internal judge/win-rate system, not an
+  external, reproducible benchmark — independent verification is
+  recommended before relying on these numbers.
+- Trained on an evolved/rewritten version of the source data, not the raw
+  human-annotated labels directly.
+- Slight general-capability regression observed post-adaptation (see table
+  above).
+- Swahili has less underlying data than Hausa/Yorùbá — performance may be
+  less stable for that language.
 
 ## Citation
 
-If you use this combined dataset, please cite all three original sources:
-
-```
-@inproceedings{muhammad2023afrisenti,
-  title={AfriSenti: A Twitter Sentiment Analysis Benchmark for African Languages},
-  author={Muhammad, Shamsuddeen Hassan and others},
-  year={2023}
-}
-
-@inproceedings{shode2023nollysenti,
-  title={NollySenti: Leveraging Transfer Learning and Machine Translation for Nigerian Movie Sentiment Classification},
-  author={Shode, Iyanuoluwa and Adelani, David Ifeoluwa and Peng, Jing and Feldman, Anna},
-  year={2023}
-}
-
-@misc{neurotech2021swahili,
-  title={Swahili Sentiment Analysis Dataset},
-  author={Neurotech-HQ},
-  year={2021},
-  howpublished={\\url{https://github.com/Neurotech-HQ/swahili-sentiment-analysis-dataset}}
-}
-```
+If you use this model, please also cite the original dataset sources
+listed in the [dataset card](https://huggingface.co/datasets/gospelgit/African-Languages_Sentiments).